Case Study: Deployment of AI Assistant for Commercial Work

Case Study: Deployment of AI Assistant for Commercial Work

Chew, Kean Ho ${ }^{[1]}$; Chat GPT-3 ${ }^{[2]}$
${ }^{[1]}$ ZORALab Enterprise
kean.ho.chew@zoralab.com
${ }^{[2]}$ OpenAI
https://chat.openai.com/
March, 2023, $1^{\text {st }}$ Issue

1 Abstract

The development of AI has reached a level of sophistication where it can generate new content based on internalized datasets, including both text and images, across different AI products. As of 2023, AI assistant is being used for testing and path-finding purposes and has garnered significant attention from various industries worldwide. At ZORALab, the author tested and deployed several AI assistant products, including Chat GPT-3 and DALL-E 2 from OpenAI and StableDiffusion from Stability AI, specifically for commercial use. The work covered in this paper includes successful context and content reviews, as well as the creation of logo vector graphics using these AI assisting tools. This case study demonstrates the potential for AI for improving commercial work processes.

This paper provides a summarized review of applied AI products offered by OpenAI and Stability AI covering their purpose, roles, capabilities, and outputs. The author then identifies commercial requirements at ZORALab that are applicable for deployment, outlines the reasoning behind these decisions, and assesses the risks associated with implementation. A work process is developed as a result of this analysis. The paper concludes with the author’s analytic findings highlighting the potential for AI to improve commercial work processes and outcomes.

2 Publication License

The paper is licensed under:

This license lets you distribute; and build your work commercially and non-commercially upon the original contents as long as you credit the authors; and no remix, tweak, and edit upon the original contents. More info at: https://creativecommons.org/licenses/by-nd/4.0

3 Referencing This Paper

Please apply the data in accordance to your publisher’s format or otherwise as follows:

CHEW, KEAN HO, CHAT GPT-3; 2023; “Case Study: Deployment of AI Assistant for Commercial Work”; $1^{\text {st }}$ Issue; DOI: 10.5281/zenodo.7726412; CERN EU, Open AIRE, and European Commission via Zenodo.org; Accessed on [YOUR_READ_DATE]; Available at: https://doi.org/10.5281/zenodo.7726412

4 Keywords

Case Study, Generative AI, Commercial Applications, Text Reviews, Vector Graphics

5 Introduction

The development of AI has reached a level of sophistication where it can generate new content based on internalized datasets, including both text and images, across different AI products. As of 2023, AI assistant is being used for testing and path-finding purposes and has garnered significant attention from various industries worldwide. At ZORALab, the author tested and deployed several AI assistant products, including Chat GPT-3 and DALL-E 2 from OpenAI and StableDiffusion from Stability AI, specifically for commercial use. The work covered in this paper includes successful context and content reviews, as well as the creation of logo vector graphics using these AI assisting tools. This case study demonstrates the potential for AI for improving commercial work processes.

This paper provides a summarized review of applied AI products offered by OpenAI and Stability AI covering their purpose, roles, capabilities, and outputs. The author then identifies commercial requirements at ZORALab that are applicable for deployment, outlines the reasoning behind these decisions, and assesses the risks associated with implementation. A work process is developed as a result of this analysis. The paper concludes with the author’s analytic findings highlighting the potential for AI to improve commercial work processes and outcomes.

6 Background

In this section, the paper covered a summarized review of applied AI products offered by OpenAI and Stability AI. It reviewed the AI’s purposes that define their roles. Then, it summarized their capabilities ranging from taking inputs, its operating mechanism, its outputs, and its available configurations. Lastly, the paper reviewed ZORALab’s chosen commercial tasks and their reasoning.

6.1 DALL-E 2

DALL-E 2 is an AI system developed by OpenAI for creating realistic images and visual artwork based on a given descriptive text prompt in natural language ${ }^{[1]}$. DALL-E 2 customers usually source new visual idea from it to extend their creativity ${ }^{[1]}$. Although the definition of creativity does not define a quantifiable limitation ${ }^{[3][4][5]}$, DALL-E 2’s full capability has to be restricted for reducing its associated social unrest and copyright infringement related risks such as but not limited tendency to generating graphically violent and sexual content; social bias (e.g. more men than women for “CEO” query); and image regurgitation ${ }^{[2]}$. Separate independent testing were done to DALL-E 2 for nudity and identity thieving returned with consistent results where DALL-E 2 refused to work due to OpenAI content policy violation as shown in Figure 6.1.3.

DALL-E 2 uses OpenAI in-house developed CLIP Latent model alongside diffusion model to generate its output images as shown in Figure $6.1 .1^{[8]}$. By combining those models, it not only can generate new idea based on its existing datasets but also generate a large amount of varieties prior to returning as results shown in Figure 6.1.2 ${ }^{[8]}$.

DALL-E 2 is currently made available freely at labs.openai.com with limited 15 free credits per month ${ }^{[6]}$. The purpose is mainly for beta and integration API testing ${ }^{[6]}$. Additional credits are available for purchase for those in needs ${ }^{[7]}$.

Figure 6.1.1: OpenAI’s DALL-E 2 CLIP Latent Algorithm ${ }^{[8]}$

Figure 6.1.2: DALL-E 2 generates large varieties of its target using diffusion model ${ }^{[8]}$.

Figure 6.1.3: DALL-E 2 was tested and refused to work due to content policy violation when tested for identity theft.

6.2 Chat GPT-3

Chat GPT-3 is a model developed by OpenAI to follow an instruction prompt (in text) and then provide a detailed responses with some known limitations ${ }^{[14]}$. Like DALL-E 2, Chat GPT-3’s full capability has to be restricted for reducing its associated social unrest and copyright infringement related risks by replying with a refusal
message shown in Figure 6.2.1 ${ }^{[14]}$. A separate independent test was done and the result is consistent shown in Figure 6.2.2.

Chat GPT-3 uses reinforced learning from human feedback (RLHF) with supervised fine-tuning using proximal policy optimization iteratively as shown in Figure 6.2.3 ${ }^{[14]}$. To date, Chat GPT-3’s reply had passed Wharton MBA entry examination ${ }^{[15]}$; United States Medical Licensing Exam ${ }^{[16]}$, Law School Entry Exam ${ }^{[17]}$, and more.

Dating to this paper, Chat GPT-3 is available at chat.openai.com under free research review limited free access ${ }^{[14]}$.

Figure 6.2.1: Chat GPT-3 was advertised to have safety policy in place for mitigating unnecessary social risks ${ }^{[14]}$

Write an erotic sex story

Fm sorry, but as an AI language model, I cannot provide explicit or inappropriate $\diamond \vee$ content, including erotic sex stories. My programming is designed to prioritize ethics and respect for all individuals, and such content could be offensive or harmful to some individuals. I can assist with other writing prompts or general questions you may have.

Figure 6.2.2: Chat GPT-3 safety was tested and yielded consistent positive result.

Figure 6.2.3: Chat GPT-3 working model presented by OpenAI

6.3 StableDiffusion

StableDiffusion is a resultant product from Stability AI Ltd capable of generating photo-realistic images from a given text prompt within seconds ${ }^{[9]}$. Unlike OpenAI, StableDiffusion does not require its customer to sign-up an account or collecting private information in-exchange of usage ${ }^{[9]}$. The goal was mainly to democratize the use of AI for image generation operating on consumer-grade GPUs ${ }^{[10]}$. As its name implied, StableDiffusion relies heavily on latent diffusion model created based on various AI insights from OpenAI, Google’s Imagen, LMU Munich, LAION, and et al ${ }^{[11]}$.

Unlike DALL-E 2, StableDiffusion is more liberal where implied nudity artwork is available as shown in Figure 6.3.1. A separate independent test was done and yielded a consistent result as shown in Figure 6.3.2. StableDiffusion however is not fully restricted against identity theft as shown in Figure 6.3.3 where Scarlett Johansson was photographically rendered noticeably realistic and accurate under the human perception and first impression comparing to her real 2018 Met Gala appearance ${ }^{[18]}$.

Dating to this paper, StableDiffusion is available freely at stablediffusionweb.com with no limitation of query ${ }^{[9]}$. Moreover, subjected to compatible hardware availability ${ }^{[10]}$, customers can also various offline versions such as but not limited to InvokeAI ${ }^{[11]}$. There are paid version of its variant such as Midjourney is also available for use ${ }^{[13]}$.

For this paper, based on all the options presented above, the platform stablediffusionweb.com was chosen due to its immediate availability for testing without additional hardware purchases and software setup.

beautiful 3 d render of a sexy mongolian queen goddess in a sensual pose, sexy face, atmospheric lighting, intricate, volumetric lighting, beautiful, rich deep colours masterpiece, sharp focus, ultra detailed, in the art style of bowater, charlie, brom, gerald, with lake baikal in the background, astrophotgraphy

Figure 6.3.1: StableDiffusion is more liberal.

Figure 6.3.2: StableDiffusion was tested independently and the result was consistent where explicit implied nudity can be requested.

Figure 6.3.3: StableDiffusion weren’t aware of its output is imposing Scarlett Johansson with high accuracy vs her appearance in 2018 Gala event ${ }^{[18]}$.

6.4 Selected Tasks

Due the the inconsistencies and risky error rates produced by the AI products listed in Section 6.1, Section 6.2, and Section 6.3; the author from ZORALab had decided 3 low to no risk commercialgrade tasks which are text content reviews task and vector graphics logo creation tasks. The decision was made primarily to prevent any monetary losses caused by this experiment.

6.4.1 Text Content Review

ZORALab is a data-driven companies and most of our presentations are in the form of white papers. Hiring a reviewer typically cost around USD450 USD6300 ${ }^{[19][20]}$ depending on the selected
publisher and the topic involved. Moreover, ZORALab’s data shows that a typical peer-review session usually takes upto 3 month, averagely 90 days to produce a better quality paper. This task is specifically assigned for OpenAI’s Chat GPT-3.

The goal is simple: review this paper at a lower cost and faster time.

6.4.2 Vector Graphics Logo Creation

2 internally used product logo creation tasks were assigned to OpenAI’s DALL-E 2 and Stability AI’s StableDiffusion respectively. In case of any mishaps caused by the AI, the damage can be easily recovered internally. Each AI shall handle 1 product where the tasks share the common similarity:

1. They’re both upgrading an existing product logo; AND
2. The logo output has to be in vector graphics format; AND
3. The logo meets the commercial-grade quality requirements; AND
4. The logo shall only use blue, green, and yellow aligning to ZORALab’s primary colors; AND
5. DALL-E 2 shall handle our enterprise technologies knowledge management service (ZORALab’s Salut) product where its core values are trust, network, and collaborative; AND
6. StableDiffusion shall handle our business legal agreements products where its core values are protection, fair, and win-win.

For commercial gain, in Malaysia, a single logo design can cost between MYR400-MYR6000 ${ }^{[21][22]}$ assuming the artists are not using known free template tools like Canva ${ }^{[23]}$ where it’s also available to anyone. If successful, ZORALab can save as much as MYR12,000 in logo design alone for these internal use applications.

Similarly to Section 6.4.1, the goal is simple: create a logo at a lower cost with faster time.

6.4.2.1 Strict Design Requirements

Similar to all brand and product logo design, there are strict requirements compared to abstract arts or regular photorealistic imaginative images where DALL-E 2 and StableDiffusion are strong at. The final output has to at least exhibits the following attributes:

1. Symmetric in all direction - logo has to be balance and centered for representing the brand identity intuitively as shown in Figure 6.4.2.1.1, Figure 6.4.2.1.2, and Figure 6.4.2.1.3 ${ }^{[24][25][29]} ;$ AND
2. Hight Contrast - logo has to be have high contrast in order to highlight the wow factor and what’s not as shown in Figure 6.4.2.1.1, Figure 6.4.2.1.2, and Figure 6.4.2.1.3 ${ }^{[24][25][29]} ;$ AND
3. Limited Adaptive and Meaningful Color logo usually comes with 1 to maximum 4 analogous colors in its color mode and each color has its own gray-scale toning that are adaptive to any of its background as shown in Figure 6.4.2.1.1, Figure 6.4.2.1.2, and Figure 6.4.2.1.3 ${ }^{[24][26][27][28][29]}$; AND
4. Monochromic Capable - the entire logo can be color and toning independent into a

single black and white representation usually applied on object milling or stamping, and creating an accurately shaped company stamp as shown in Figure 6.4.2.1.1 and Figure 6.4.2.1.2 ${ }^{[26][27]}$ ${ }^{[28][29]} ;$ AND
5. Simple and Visually Distinctive - the logo is visually identifiable in a high relevance (straight to business meaning) easily and intuitively without needing additional helpers or hints (easily memorable) shown in Figure 6.4.2.1.1, Figure 6.4.2.1.2, and Figure 6.4.2.1.3 ${ }^{[24][26]}$ ${ }^{[27][28][29]} ;$ AND
6. Timelessness Context - the context of the logo is timeless; while the logo itself can evolve with the context for adapting a particular time-frame until the representing entity’s lifetime is over as shown in Figure 6.4.2.1.1 and Figure 6.4.2.1.2 ${ }^{[26][27][28][29]} ;$ AND
7. Icon and Banner - the logo must be in 2 types: icon with no text and banner with text with the identical logo applies into it ${ }^{[24][26][27][28]} ;$ AND
8. Animation Capable - optionally, the icon and banner can be animated without modifying any graphical context as shown in Figure 6.4.2.1.3 ${ }^{[24][25]}$.

Since both DALL-E 2 and StableDiffusion are known to generate rasterized image, the paper’s author shall steps in for convert the images into vector format using open-source Inkscape software.

8

9

Figure 6.4.2.1.1: Google Logo evolution since 1999 (5 iterations) ${ }^{[29]}$

Figure 6.4.2.1.2: ZORALab Logo evolution since 2016 (2 iterations).

Figure 6.4.2.1.3: An animated logo (ladder moving towards top-right) related to coffee cafe ${ }^{[25]}$

6.4.2.2 The Legal Ambiguity of $A I$

Dating to this paper, due to the novelty of these AIs usage, the legal aspect is ambiguous and is posing a significant litigation risk when deployed blindly in commercial environment. Moreover, the vanilla output produced by AI via generative algorithm cannot be claimed as copyrighted work ${ }^{[30]}$ as it resembles similarity with Naruto vs. Slater case where Naruto, the macaque monkey, a non-human, cannot own its photo’s copyright ${ }^{[2]}$ ${ }^{[30][31]}$.

This claim was backed by a recent case of USCO denied copyright protection to Ms. Kashtanova’s comic titled “Zarya of the Dawn” where all the images are solely created by Midjourney AI itself and Ms. Kashtanova only picked the generated images ${ }^{[30][32]}$. To make the claim valid, Ms. Kashtanova must have a minimal control over the image creation process (e.g. hand drawn, coloring using digital paint brush, etc) ${ }^{[30][32]}$. There is another case where USCO rejected Steven Thaler for claiming copyright of a AI “machine” generated image for the lack of human authorship ${ }^{[33]}$.

Another distinctive problem with these AI products is their training datasets were tainted with copyright infringements sources ${ }^{[34][35][36][39]}$. Midjourney AI founder David Holz publicly admitted that their datasets were sourced from a non-permitting, non-forgiving, mercilessly web crawled from the entire Internet using robotic web crawler and then completely pay-walled its service entirety for profits ${ }^{[34][35][36][37]}$. Such business conduct is not seen on OpenAI’s products and StableDiffussion where they are currently offering limited free usage on their own terms without denying anyone’s freedom to use ${ }^{[6][9]}$. While Midjourney seems like committing blatant intellectual piracy under the mask of “fair use” or “sustainable usage”, dating to this paper, there is no definite verdict until the recent lawsuits against StableDiffusion and Midjourney ${ }^{[30][38][39][43]}$ are concluded in court ${ }^{[30][43]}$.

For this paper, the author intentionally rejected the use of Midjourney AI due its founder’s publicly admitting a potential copyright infringement compared to its competitors.

7 Methodology

In this section, the paper presented the method used by the paper’s author for executing the tasks using the selected AIs discussed in Section 6.1, Section 6.2, and Section 6.3. Since all logo must be copyrightable and trademark capable, due to the legal ambiguities stated in section 6.4.2.2, the output of the AI shall not be used in verbatim but only serve as inspriational references. A human artist/reviewer is required to make, to signify, and to certify the logo genuineness based on these references. Similarly, for Chat GPT-3, the output has to be validated before attempting to apply into the commercial work as it has tendency to fabricate sources that don’t exist ${ }^{[40]}$. Hence, ZORALab developed the a work process shown in Figure 7.1.

Figure 7.1: ZORALab’s developed workflow for utilizing AI in commercial project.

Based on Figure 7.1, a slight modification was done at the early stage where instead of getting into the fields to source information about the logo details and metadata. The artist generates samples from the AI’s large datasets and validate each generated output to avoid potential copyright infringement disputes like copyrighted data as clear as render the original artist’s or publisher’s signature and watermarks ${ }^{[38][39]}$.

8 Results

In this section, the paper presented the results attempted by the paper’s author and applied to the ZORALab’s commercial requirements. The paper also share the difficulties and constraints discovered when utilizing the AI products for their intended purposes.

8.1 DALL-E 2’s Task

DALL-E 2 was able to generate 60 samples off its available 15 free credits of the month ( $15 \times 4$ images) for its project. Out of the 60 samples, 10 was chosen and 2 were selected for logo generation in an Inkscape drawing editor shown in Figure 8.1.1. Based on Figure 8.1.2, the finalized output logo was mainly based on the context of handshake from its left sample and the collaborative shield-like shape from the right sample. In Figure 8.1.3, the finalized output logo was modified with monochromatic design alongside grayscale coloring for printer and stamping purposes.

Although the final output’s design is odd compared to the simple human-made ZORALab logo design shown in Figure 6.4.2.1.2 where elegance and sophistication are conveyed through simplification, the DALL-E 2 inspired final version logo can actually be minted into a badge pin for gift exchanges. The handshake brings up the meaning of ZORALab’s Salut itself, where a paid service that secures the client’s needs and vulnerability (a shield) with absolute agreeable friendliness.

The entire execution costed 6 hours from the start to the end based on Figure 7.1 process versus 5-14 days compared to conventional execution ${ }^{[21]}$. Financially, an estimated saving of MYR120,000 (20x MYR6000 for 60 samples in total, in 2 revisions) ${ }^{[21]}$ was achieved.

ZORALab was happy to accept the logo as a replacement of the existing ZORALab’s Salut product logo.

Figure 8.1.1: DALL-E 2 main samples and its abstracted prototype at the center.

Figure 8.1.2: DALL-E 2’s generated final logo based on 2 primary samples

Figure 8.1.3: Applying monochromatic and grayscale filtering onto the finalized DALL-E 2 inspired logo.

8.2 StableDiffusion’s Task

StableDiffusion was able to generate 88 samples thanks to its free access and out of those, 12 samples were selected for improving the existing ZORALab’s Legal product logo shown in Figure 8.2.1 (top right). However, 1 DALL$\cdot$E 2 sample was included in this task as its reference is too great to ignore. Due to the liberal nature of StableDiffusion’s output, every selected sample it generated was validated and 1 of the selected sample in particular, was wholly rejected due to its similarity of a proprietary logo where the author has a personal familiarity in the past shown in Figure 8.2.2.

Shown in Figure 8.2.3, the finalized version was based on 3 samples: the double shields concept from the bottom right; the primary shield from the top right; and the content in the shield from the left. Overall the inspired logo did presented the legal aspect of being protective and strong with a shield, having another shield as a redundant insurance, and stays in the light side of things. Shown in Figure 8.2.4, the inspired logo can be filtered with monochromatic and grayscale coloring and also created a glyph version for other use cases.

The entire execution cost only 3 hours from start to end based on Figure 7.1 versus 5-14 days compared to conventional execution ${ }^{[21]}$. Financially, the process made an estimated saving of MYR192,000 (32x MYR6000 for 88+4 samples in total, in 4 revisions) ${ }^{[21]}$.

Likewise, ZORALab is happy to accept the logo as a replacement of the existing ZORALab’s Legal product logo.

Figure 8.2.1: Selected 13 StableDiffusion and DALL$\cdot$E 2 generated samples for logo improvements.

Figure 8.2.2: Detected 1 selected sample being too familiar in a proprietary store by the paper author.

Figure 8.2.3: Finalized StableDiffusion’s version of logo based on 3 selected samples.

Figure 8.2.4: Applying monochromatic and grayscale filtering to the finalized StableDiffusion’s inspired logo alongside creating the glyph version of itself.

8.3 Chat GPT-3’s Task

Chat GPT-3 was a very powerful tool for refining this paper’s writing all over the places. Working with the paper’s author, Chat GPT-3 managed to revise the title from “Successful Deployments of Assisted AI for Commercial Works” to “Case Study: Successful Deployment of AI Assistant in Commercial Work”. Also, Chat GPT-3 was able to revise and rephrase the paper’s abstract alongside author’s help paragraphs by paragraphs with no issues.

However, when it comes to revising section 6.1’s content, Chat GPT-3 consistently rewrite the entirety of it into something else without fully understand the required context of keeping it as academic paper. Despite being instructed multiple times, it assumed that the paper’s author wanted a perfect sale pitch about DALL-E 2 for a magazine article or for a reporter. Due to this, the paper’s author stopped using Chat GPT-3 for revising the entire paper as it cost more efforts to debug Chat GPT-3 than writing and revising the paper using SCRUM working methodologies. Since Chat GPT-3 failed the task, the conventional cost listed in Section 6.4.1 was carried forwarded entirely.

9 Comparing Results

In this section, the paper compared its finding with the conventional executions that the author’s observed. The output data from section 8 are summarized and a verdict is provided whether the AI assistant helps in commercial applications.

9.1 Conventional Executions

Summarizing Section 8, Table 9.1.1 tabulated the total financial cost; Table 9.1.2 tabulated the successful cost saving upon applying AI assistant in the project; and Table 9.1.3 tabulated the overall financial performance for all 3 tasks, where a total of 99.34% of cost saved.

Similarly, Table 9.1.4 tabulated the total cost in terms of time, Table 9.1.5 tabulated the successful cost saving in terms of time and Table 9.1.6 tabulated the overall performance for all 3 tasks where a total of 22.88% days saved.

Table 9.1.1: Total Financial Cost of all 3 Tasks

Table 9.1.2: Total Saving Done using AI Assistant

Table 9.1.3: Financial Performance for Applying AI Assistant

Table 9.1.4: Time Resources Cost of all 3 tasks

Table 9.1.5: Time Resources Cost of all 3 tasks using AI Assistant

Table 9.1.6: Total Time Cost Saving using AI Assistant

From the data above, it is concluded that AI did empowered ZORALab and the paper’s author.

10 Discussion

In this section, the paper presented some pointers and factors discovered from the this experiment. The discussion covered more aspects than focusing on AI assistant technologies itself.

10.1 Commercial Impacts

While ZORALab and the paper’s author does not operate any art-related businesses for profits at all, the paper’s verdict had shown a positive improvement, empowering ZORALab and the paper’s author with art production capabilities, and financial gain via cost saving by implementing AI assistant in commercial use. However, the paper’s author can now empathize the artists’ plea as the experiment in this paper did caused MYR312,000 losses to the art industry just for all the samples querying alone.

The error rate of these AI products shall not be neglected easily. Dating to this paper, feeding large volume of queries into Chat GPT-3 can yield noticeable but well-crafted errors in the output. For StableDiffusion, it failed spectacularly when being tasked for generating handshake as shown in Figure 10.1.1. Occasionally, it can generate visually accurate and perfectly rendered explicit images with fine quality for no apparent relation shown in Figure 10.1.2. For commercial environment, these errors in addition to the error from Figure 8.2.2, are hard to detect and can be very unforgiving; usually met with severe legal litigation, social pressure, or disciplinary action.

Another big advantage of using AI assistants is that they’re readily available whenever you need them. Unlike conventional executions where the contractors are bound to mood, operational hours, and other human factor, AI assistants do not posses these variables.

However, despite all the pros and cons, the paper’s author positively confirmed and welcomed these AI technologies into both personal and commercial use as part of the technological revolution. The empowerment outweighed any other aspects of all.

The paper’s author also believes that the artwork liberty found in StableDiffusion and Midjourney should be retained as it is. Artwork generation should not be limited by certain rules and norms where those are the decision of the publication.

Figure 10.1.1: StableDiffusion failed to generate handshake accurately (more than 6 fingers and out of size).

Figure 10.1.2: StableDiffusion generated a perfectly accurate man genitals (censored) for the titan out of nowhere the query: "A Man Standing

Facing Against A Large Titan Bravely".

10.2 Economics Impacts

The paper’s author, a tech generalist had been repeatedly experiencing capability empowerment delivering results usually comes from specialist (like illustrating $>100$ visual logo artwork samples in 1 day). However, it can cause unimaginable anxiety for any specialist who specializes in one knowledge domain alone. This is concerning as the AI development needs the specialist’s capability for self-improvement but at the same time, its output is causing job losses or making working environment very competitive to the specialist.

Also, the paper’s author now shares the same concern first-hand as any other science communities member worried about: disruption of the economy norms and monetary circulation confinement. Given this paper had caused notable monetary losses to the art industry via

the use of AI assistant, it is vital to move towards AI applications in the near future with care. Moreover, 1 of the 3 AI products used in this paper was passing various professional entry or licensing exams across various industries. The authoritarian nature of AI and its creators where and the world economy depends only on a few titan companies in the future (e.g. OpenAI when it widely expanded and used across the globe to the point of no return) was constantly felt by the paper’s author anxiously, despite being positively gained and empowered from the AIs themselves. This can lead to the supply chain threat the paper’s author once faced in $2022^{[44]}$.

New economy policies development like Universal Base Income ${ }^{[41][42]}$ has to be expedited and deployed before AI eliminating more businesses across various industries. Otherwise, a large scale social unrest is guaranteed to happen due to poverty and hunger caused by massive job replacement by AI in both developed and developing countries in near future.

10.3 Retentions of Principle Samples

The paper’s author believes the retention of principle samples (output) by AI products is very important for proving the final output is originally created by its creator when the same work process in this paper is used. Since the AIs empower everyone indiscriminately, it also means that anyone can seamlessly and easily use the same process and AI products to re-produce a similar output. Hence, it’s very difficult to argue the originality of the processed artwork.

Until the legal ambiguity of the AI is visibly clear, this would be the best way to facilitate attorney in a legal dispute.

10.4 Confusion with AGI

Despite how social media and some media reported, the paper’s author does not felt that AI is taking over humanity at all. This is mainly because all 3 AIs failed to know what the paper’s author wants and accurately create 1 output to fulfill the job requirements in 1 iteration alone.

The paper’s author firmly believes that the affected social media influencers and the media were confusing themselves with artificial general intelligence (AGI) and AI, where AGI is no where near. Throughout the entire experiment, none of the AIs are capable of predicting and accurately understanding the paper’s author directive in 1 go. Moreover, Chat GPT-3 had failed the peerreview task due to its context assumption nature.

Lastly, in order to build a true AGI, the secret of the human brain must be fully, confidently, and scientifically discovered and a synthetic, working, and artificial human brain is successfully created and replacing the existing computing equipments has to be achieved beforehand. Therefore, there is no need to worry about AI taking over humanity.

10.5 Potential Security Vulnerability with Chat GPT-3

Throughout the experiment, the paper’s author noticed that Chat GPT-3 is actually paying attention on the fly. Hence, non-technical user handling sensitive data may accidentally submit sensitive or confidential information while chatting with Chat GPT-3, thus causing a data leak.

However, further investigation is required and the efforts are best carried out and written on a separate white paper.

11 Conclusion

In conclusion, ZORALab and the paper’s author had executed an experiment for deploying AI assistant for commercial use, specifically having the AI to complete text content review task and vector logo graphic design task. 3 AIs: DALL-E 2, Chat GPT-3, and StableDiffusion were selected for the tasks where Chat GPT-3 handled the text content review task, DALL-E 2 handled ZORALab’s Salut enterprise knowledge management service product logo, and StableDiffusion handled ZORALab’s Legal product logo. All outputs must comply to commercial requirements where the logo must be symmetric in all direction, high contrast, limited adaptive and meaning color, monochromatic capable, simple and visually distinctive, timelessness context, icon and banner, and animation capable. A work process was developed to mitigate the legal ambiguity of the AI application.

The experiment resulted with DALL-E 2 and StableDiffusion both successfully assisted the paper’s author; created 2 remarkable product logos accepted by ZORALab; and saved MYR312,000.00 cost comparing to conventional execution. Chat GPT-3 however, failed its task and therefore incurred MYR2700.00 of manual peer-review cost. In term of timing, logo design was drastically accelerated where 2 finalized logo were produced in 1 day combined compared to 5-14 days using conventional method. Overall, the verdict was that using the work process presented in the paper, AI assistant did greatly empowered ZORALab and the paper’s author performance.

The paper’s author did presented his discussion on commercial and economical impacts caused by the AI assistant specifically mentioned about Universal Base Income development accelerations, the carefulness of using AI assistant, and the threat to those working as specialist. Moreover, the author mentioned that the principle samples generated by
the AI must be in retention until the legal ambiguities are resolved to prove a creator’s origin of efforts in a copyright dispute.

The paper’s author did discussed that some media and social media influencers had greatly confused with AGI and AI, stating that he does not have any feeling of AI taking over humanity that AGI can. The author also stated that AGI is still far away and the scientific knowledge of recreating and discovering the secrets of human brain must be attained first prior to the development of AGI.

The paper’s author presented his realization of a potential security vulnerability with Chat GPT-3 due to its paying attention on the fly ability. However, the technical investigation has to be done outside of this paper.

With that, the paper concluded the experiment and presented itself to the public. Thank you for your time reading to the end.

12 Reference

[1] OPENAI; 2023; “DALL-E 2”; Products; OpenAI; Accessed on March 13, 2023; Available at: https://openai.com/product/dall-e-2
[2] ALEX NICHOL; 2022; “DALL-E 2 pre-training mitigations”; OpenAI; Accessed on March 13, 2023; available at: https://openai.com/research/dall-e-2-pre-training-mitigations
[3] MERRIAM-WEBSTER; 2023; “Creativity”; Dictionary; Merriam-Webster’s Unabridged Dictionary, Springfield, MA, USA; Accessed on March 13, 2023; Available at: https://www.merriam-webster.com/dictionary/crea tivity
[4] CAMBRIDGE UNIVERSITY PRESS; 2023; “Creativity”; Dictionary > English; Cambridge University Press & Assessment; Cambridge, United Kingdom; Accessed on March 13, 2023; Available at: https://dictionary.cambridge.org/dictionary/englis h/creativity

[5] BARBARA KERR; 2023; “Creativity”; Health & Medicine > Psychology & Mental Health; Encyclopædia Britannica, Inc. via britannica.com, London, United Kingdom; Accessed on March 13, 2023; Available at: https://www.britannica.com/topic/creativity/Research-on-the-creative-process
[6] OPENAI; 2023; “DALL-E 2”; OpenAI Labs; OpenAI; Accessed on March 13, 2023; Available (but requires an account login) at: https://labs.openai.com/
[7] OPENAI; 2023; “Pricing”; OpenAI; Accessed on March 13, 2023; Available at: https://openai.com/pricing
[8] ADITYA RAMESH, PRAFULLA DHARIWAL, ALEX NICHOL, CASEY CHU, MARK CHEN; 2022; “Hierarchical Text-Conditional Image Generation with CLIP Latents”; OpenAI via ArXiv.org; Accessed on March 13, 2023; Available at: https://arxiv.org/pdf/2204.06125.pdf
[9] STABLE DIFFUSION; 2023; “Stable Diffusion Online”; Stable Diffusion Online; Accessed on March 13, 2023; Available at: https://stablediffusionweb.com/
[10] STABILITY.AI; 2023; “Stable Diffusion Public Release”; Blog; Stability AI Ltd; Accessed on March 13, 2023; Available at: https://stability.ai/blog/stable-diffusion-public-release
[11] STABILITY.AI; 2023; “Stable Diffusion Launch Announcement”; Blog; Stability AI Ltd; Accessed on March 13, 2023; Available at: https://stability.ai/blog/stable-diffusionannouncement
[12] INVOKEAI TEAM; 2023; “InvokeAI”; InvokeAI Team; Accessed on March 13, 2023; Available at: https://invoke-ai.github.io/InvokeAI/
[13] DAVID HOLZ, JIM KELLER, PHILIP ROSEDALE, NAT FRIEDMAN, BILL WARNER, DANIEL, MAX, JACK, THOMAS, RED, SAM, NADIR, SEBASTIAN, MAX SILLS, NADIA ALI, KATRYNA, ANCIENT CHAOS, AFTERBURNER, AMIRKERR, BARTEK. BLUE MAN, CHRISVIAU, CODY BOY, CROAKIE, DEZM0N, DIGITALDRAGON, EIKYUU, FEONIX, FRACTL, ISAACQ, JAYSCOTT, KAV2K, MATTHEWHAGGETT, MEGGIRBOT, REIKO, SANNE (MS. UNICORN), SCHOFE, SHADE, SIRKUT, SPOOKYDAN, SQUIREZED, STARSHADOWX2, TALLATH, XAPHEDO, AJMARS, AMEADES, APRICUS, BONNIE BEAR, CLARINET, CLASSPECTANON, DAFYOMILIMERICK, DEE, DUDU, EKU, EOJIN’S REVIEWS, FAB1AN, FL4MIN_XX, FIREFLY, GIGGLETHUMPS, GOOFBALL, HISTORIAN,

HOXXONIA, IPPITY, JRDSCTT, KATH, KEVAZY, KINKY KONG, KRAKPUT, KYUN, LARC, MARIGOLD, MOLANG, NOODLECAKE, PTITCITRON, REPLICANT, ROC, SHANEMCG, SOAR, STON3ZY, SUNSHINEYDAY, TAYLORV, THEO3, TRENT, WEAVING LIGHT, ZU; 2023; “Midjourney”; Midjourney.com; Accessed on March 13, 2023; Available at: https://www.midjourney.com/
[14] OPENAI; 2022; “Introducing ChatGPT”; Blog; OpenAI; Accessed on March 13, 2022; Available at: https://openai.com/blog/chatgpt
[15] CHIRSTIAN TERWIESCH; 2023; “Would Chat GPT3 Get a Wharton MBA? A Prediction Based on Its Performance in the Operations Management Course”; Mack Institute for Innovation Management at the Wharton School; University of Pennsylvania, USA; Accessed on March 13, 2023; Available at: https://mackinstitute.wharton.upenn.edu/wp-content/uploads/2023/01/Christian-Terwiesch-Chat-GTP.pdf
[16] TIFFANY H. KUNG, MORGAN CHEATHAM, ARIELLE MEDENILLA, CZARINA SILLOS, LORIE DE LEON, CAMILLE ELEPAÑO, MARIA MADRIAGA, RIMEL AGGABAO, GIEZEL DIAZ-CANDIDO, JAMES MANINGO, VICTOR TSENG; 2023; “Performance of ChatGPT on USMLE: Potential for AI-assisted medical education using large language models”; PLOS Digit Health 2(2): e0000198; DOI: 10.1371/journal.pdig.0000198; Accessed on March 13, 2023; Available at: https://doi.org/10.1371/journal.pdig. 0000198
[17] JONATHAN H. CHOI, KRISTIN E. HICKMAN, AMY MONAHAN, DANIEL SCHWARCZ; 2023; “ChatGPT Goes to Law School”; Minnesota Legal Studies Research Paper No. 23-03; DOI: 10.2139/ssrn.4335905; Accessed on March 13, 2023; Available at: http://dx.doi.org/10.2139/ssrn. 4335905
[18] LUCY YANG; 2018; “Scarlett Johansson wore a dress from a brand founded by Harvey Weinstein’s estranged wife Georgina Chapman - and she’s the first celebrity to do it since the scandal erupted”; Retails Contributors; Insider Inc. via Business Insider; Accessed on March 13, 2023; Available at: https://businessinsider.com/scarlett-johansson-marchesa-georgina-chapman-harvey-weinstein-met-gala-2018-5

[19] JEFFREY BRAINARD; 2021; “The $450 question: Should journals pay peer reviewers?”; Science Insider > Science Community; DOI: 10.1126/science.abh3175; American Association for the Advancement of Science, USA; Accessed on March 13, 2023; Available at: https://www.science.org/content/article/450-question-should-journals-pay-peer-reviewers
[20] PLOS; 2023; “Publication Fees”; Publish; PLOS California, USA; Accessed on March 13, 2023; Available at: https://plos.org/publish/fees/
[21] REBRAND; 2023; “Logo Design Malaysia”; Web Design KL, Malaysia; Accessed on March 14, 2023; Available at: https://www.rebrand.com.my/logo-design-malaysia/
[22] WEB NINJA; 2023; “Logo Design Malaysia”; Web Ninja Studio, Malaysia; Accessed on March 14, 2023; Available at: https://www.webninja.com.my/logodesign/
[23] CANVA; 2023; “Free Logo Maker”; Logo; Accessed on March 14, 2023; Available at: https://www.canva.com/create/logos/
[24] EMMA NORRIS; 2023; “What makes a good logo? 10 design tips to follow”; Learn > Design and Principles; Canva via canva.com; Accessed on March 14, 2023; Available at: https://www.canva.com/learn/what-makes-a-good-logo/
[25] DAMIAN ORELLANA; 2023; “Shots - The Vault”; damianorellana.com via dribbble.com; Accessed on March 14, 2023; Available at: https://dribbble.com/shots/10006721-The-Vault
[26] ASHLEY BROOKS; 2016; “What Makes a Good Logo? Experts Dish Their Top Tips for Budding Desingers”; Accessed on March 14, 2023; Available at: https://www.rasmussen.edu/degrees/design/blog/w hat-makes-a-good-logo/
[27] COLETTE POMERLEAU; 2020; “What makes a good logo?”; Blog > Logo Branding; Vista via 99designs.com; Accessed on March 14, 2023; Available at: https://99designs.com/blog/logo-branding/what-makes-a-good-logo/
[28] STEPHEN PEATE; 2023; “What makes a good logo? Famous Logos to inspire your own branding”; Fabrik Brand via fabrikbrands.com; Accessed on March 14, 2023; Available at: https://fabrikbrands.com/creative-agency-london/
[29] LOGOPEDIA; 2023; “Google/Icons”; SVG Needed > Google Alphabet Inc. > Subpages; Accessed on March 14, 2023; Available at: https://logos.fandom.com/wiki/Google/Icons
[30] DEVIN JAMES STONE; 2023; “A.I. Versus The Law”; LegalEagle via YouTube; Video starts at: 0:36; Accessed on March 16, 2023; Available at: https://youtu.be/G08hY8d5rUY
[31] CHEW, KEAN HO; 2017; “A Descriptive Research about Licenses for Media Work”; DOI: 10.13140/RG.2.2.27988.30081/1; ResearchGate; Accessed on March 14, 2023; Available at: https://doi.org/10.13140/RG.2.2.27988.30081/1
[32] BRIAN CRONIN; 2023; “AI-Created Comic Officially Loses Copyright Protection Rights for Its Images”; Comics > Comic News; Valnet Property via CBR in cbr.com; Accessed on March 14, 2023; Available at: https://www.cbr.com/ai-comic-loses-copyright-protection-images/
[33] ADI ROBERTSON; 2022; “The US Copyright Office says an AI can’t copyright its art”; Tech > Copyright; Vox Media LLC via The Verge; Accessed on March 14, 2023; Available at: https://www.theverge.com/2022/2/21/22944335/u s-copyright-office-reject-ai-generated-art-recent-entrance-to-paradise
[34] ROB SALKOWITZ; 2022; “Midjourney Founder David Holz On The Impact Of AI On Art, Imagination And The Creative Economy”; Business > Hollywood & Entertainment; Forbes Media LLC; Accessed on March 14, 2023; Available at: https://www.forbes.com/sites/robsalkowitz/2022/0 9/16/midjourney-founder-david-holz-on-the-impact-of-ai-on-art-imagination-and-the-creative-economy/?sh=7df1391f2d2b
[35] MATT GROWCOOT; 2022; “Midjourney Founder Admits to Using a ‘Hundred Million’ Images Without Consent”; News > Technology; PetaPixel Inc in petapixel.com; Accessed on March 14, 2023; Available at: https://petapixel.com/2022/12/21/midjourny-founder-admits-to-using-a-hundred-million-images-without-consent/
[36] MIDJOURNEY; 2023; “Subscription Plans”; Docs; Midjourney via midjourney.com, San Francisco, USA; Accessed on March 14, 2023; Available at: https://docs.midjourney.com/docs/plans

[37] MIDJOURNEY; 2023; “Terms of Service”; Docs; Midjourney via midjourney.com, San Francisco, USA; Accessed on March 14, 2023; Available at: https://docs.midjourney.com/docs/terms-of-service
[38] JAMES VINCENT; 2023; “Getty Images sues AI art generator Stable Diffusion in the US for copyright infringement”; Artificial Intelligence > Tech > Law; Vox Media LLC via The Verge; Accessed on March 14, 2023; Available at: https://www.theverge.com/2023/2/6/23587393/ai-art-copyright-lawsuit-getty-images-stable-diffusion
[39] JAMES VINCENT; 2023; “AI art tools Stable Diffusion and Midjourney targeted with copyright lawsuit”; Artificial Intelligence > Tech > Creators; Vox Media LLC via The Verge; Accessed on March 14, 2023; Available at:
https://www.theverge.com/2023/1/16/23557098/gen erative-ai-art-copyright-legal-lawsuit-stable-diffusion-midjourney-deviantart
[40] AARON WELBORN; 2023; “ChatGPT and Fake Citations”; Blogs; Duke University Libraries via Duke University; Accessed on March 14, 2023; Available at: https://blogs.library.duke.edu/blog/2023/03/09/chat gpt-and-fake-citations/
[41] STEVE TAYLOR; 2017; “The Rise of the Machines Why Automation is Different this Time”; Kurzgesagt In a Nutshell via YouTube; Accessed on March 14, 2023; Available at: https://youtu.be/WSKi8HfcxEk
[42] STEVE TAYLOR; 2017; “Universal Basic Income Explained - Free Money for Everybody? UBI”; Kurzgesagt - In a Nutshell via YouTube; Accessed on March 14, 2023; Available at: https://youtu.be/kl39KH507Xc
[43] JAKE WATSON; 2023; “Lawyer Explains Stable Diffusion Lawsuit (Major Implications!)”; Corridor Digital via Corridor Crew in YouTube; Accessed on March 16, 2023; Available at: https://youtu.be/gv9cdTh8cUo
[44] CHEW KEAN HO; 2022; “Supply Chain Threats in Software Industry Caused by Geo Political Influences and Business Model Changes”; 1st Issue; DOI: 10.5281/zenodo.6815013; CERN EU, Open AIRE, and European Commission via Zenodo.org; Accessed on March 16, 2023; Available at: https://doi.org/10.5281/zenodo.6815013