ChatGPT
ChatGPT , for Chat Generative Pre-trained Transformer , is a prototype of a conversational agent ( chatbot ) using artificial intelligence , developed by OpenAI and specialized in dialogue. It is based on OpenAI's major language models GPT-3.5 and GPT-4 .
ChatGPT is able to answer questions, hold conversations, generate computer code, and write, translate or synthesize texts. It can do this by taking into account the context, and constraints such as writing style. The paid versions (ChatGPT “Plus”, “Team” and “Enterprise”) provide access to the more advanced GPT-4 model as well as specialized chatbots, and allow analysis and generation of images.
Due to its multiple capabilities, the prototype raises concerns about the risks of misuse for malicious purposes, plagiarism in academia and job cuts in certain sectors. ChatGPT also raises security and misinformation concerns , as the model can be used to create fake text and misleading information.
Launched inin a free, non-Internet version, ChatGPT enjoys wide media exposure and receives generally positive reception, although its factual accuracy is criticized. In, ChatGPT has more than 100 million registered accounts, and the OpenAI company is then valued at 29 billion US dollars 2
Name
The acronym ChatGPT is a portmanteau made up of the English words “ chat ” and “ GPT ” 3 .
The word “ chat ” refers to an online dialogue in which Internet users exchange messages instantly. The particularity of ChatGPT is that it allows an Internet user to chat not with other Internet users but with a system based on artificial intelligence 4 .
The word “GPT” is an acronym meaning “ Generative Pre-trained Transformer ” 4 .
Features
ChatGPT is an artificial intelligence conversational agent or “ chatbot ”, in other words a virtual assistant that uses artificial intelligence to communicate with its users 5 .
The conversational robot is available in multiple languages, including French , and offers variable performance depending on the language 6 , 7 . It remembers previous messages given to it by the user during the same conversation 8 . It is capable of answering test questions in a language very close to that of a human, or even, depending on the question, with a higher level of performance than an average human respondent 9 , 10 .
It also has generative capabilities to produce tailor-made textual content. The robot can in particular generate articles, essays or poems on different tones and subjects 11 . It is also used to produce and correct computer code 12 .
Access to ChatGPT is free, but requires creating an account on the OpenAI website 13 . Users help train the robot through their queries and their evaluation of responses 14 .
The ChatGPT mobile application, launched in May 2023 on iOS 15 and then on Android 16 , integrates a voice recognition technology called Whisper which allows you to converse by voice with the robot 17 .
ChatGPT has been trained to reduce the rate of false or harmful responses. Furthermore, filtering is carried out by a moderation API , messages classified for example as racist or sexist being rejected 18 , 8 .
Functioning
ChatGPT integrates the GPT-3.5 and GPT-4 pre-trained generative transformers in a conversational format. The “pre-training” of GPT-3.5 and GPT-4 consists of predicting, given a part of a text, the following token (a token being a piece of text, typically a word, part of a word , or punctuation 19 ). This training in predicting what will follow, repeated for a large number of texts, allows these models to accumulate knowledge. They can then generate text similar to that used for pre-training, by predicting the following tokens one by one 20 , 21 . In the case of GPT-4 Vision, training was also performed on images, making the model capable of analyzing images 22 .
For the second learning phase, humans wrote queries, and the corresponding expected responses. ChatGPT is trained to provide similar responses, which pushes it to adopt a conversational format where it plays an assistant role 18 .
A third learning phase uses a reinforcement learning technique called RLHF to make the model more truthful, useful and harmless. A “reward model” was first trained to recognize responses that humans approved or disapproved of. ChatGPT was then trained to provide responses that satisfied this model representing human preferences 21 , 18 .
After deployment, ChatGPT continues to be regularly trained, notably based on how users judge responses (thumbs up or down) 23 .
When ChatGPT responds to a user, the text of the conversation up to that point is first split into tokens, and each token is converted into a vector that encodes its meaning 24 and position in the text. Inside a transformer like GPT-3.5 or GPT-4 is a succession of attention layers and short neural networks . Layers of attention combine concepts together, allowing context and complex relationships to be taken into account 25 . At the output, the transformer assigns a score to each possible token, and a token is selected partially randomly among those with the highest score. ChatGPT thus generates the tokens one by one, until generating a special token indicating the end of the response 26 .
Boundaries
ChatGPT still has certain limitations:
- Knowledge end date : The data used for training ChatGPT stops on a certain date, and ChatGPT is basically unaware of events that occur afterward. This date was previously set for September 2021, but is now January 2022 for ChatGPT 3.5 and April 2023 for the paid version ChatGPT 4 27 . Only the paid version ChatGPT 4 can decide to use the internet if it needs more recent information 28 .
- Unstable quality : The quality of ChatGPT responses fluctuates and does not necessarily increase from one version to another 29 .
- Bias : ChatGPT may reproduce biases inherited from its training data, for example biases related to gender 30 or politics 31 .
- Hallucinations : ChatGPT can generate information that appears plausible but is false. This phenomenon was particularly marked in the first versions 32 , 33 .
Use
THE, OpenAI CEO Sam Altman indicates that the prototype, which was then free, reached one million users 34 .
In, ChatGPT exceeds 100 million registered accounts, two months after its launch, and in March recorded 1.6 billion visits 35 , making it the fastest growing application then 36 .
According to François Fleuret, professor at the University of Geneva , the profile of ChatGPT users is quite varied. These may be students , teachers , developers needing snippets of code from a computer program or even professionals wishing to have text suggestions for writing sensitive emails 37 .
The ChatGPT 4 version allows online search through Bing integration . This feature is disabled ondue to copyright because it allowed the entire text of web pages to be displayed, some of which were behind paywalls 38 , but was reactivated at the endin the paid version, which takes into consideration the reading prohibitions entered in the robots.txtweb page file 39 .
Integration into Microsoft solutions
Microsoft maintains a close partnership with OpenAI, having invested nearly $13 billion in OpenAI 40 (and thus owning 49% of the company's capital 41 ). OpenAI has granted Microsoft exclusivity for the commercial exploitation in the cloud of its generative AI models such as GPT-4 40 .
Microsoft Bing
After a test phase reserved for a limited number of users, Bing 's conversational agent called "Microsoft Copilot" (initially "Bing Chat") and based on GPT-4 becomes accessible to all in May 2023 42 . The latter can provide links to sources that he used to produce his answer.
It is present in both interfaces of the new version of Bing:
- in the classic search interface, it appears in a box when the user enters a query in the form of a question. It then provides a synthetic response complementary to the list of websites provided by the engine 43 ;
- a new interface entirely dedicated to conversation is also created. In this, there is no longer a list of links and the Internet user asks his questions directly to the robot in order to obtain synthetic results without having to search for them himself on the web 43 .
The objective for Microsoft would thus be to allow its Microsoft Bing search engine to compete with Google, which holds more than 90% of the global market share in 2023 44 .
Microsoft Edge
ChatGPT has also been integrated into the Microsoft Edge web browser, allowing it to comment on PDF files , summarize them, add information from the web or even translate them 45 .
Microsoft Office
Microsoft also integrated GPT-4 in 2023 in a solution called "Microsoft 365 Copilot", which is used in Microsoft Word , Excel , Outlook and Teams and PowerPoint 46 , 47 . In Microsoft Word, this can be used to write, edit or summarize text, and to create images or graphics. Use cases also include the creation of meeting minutes, the synthesis of documents, or assistance in drafting emails 46 .
Economic model
The version of ChatGPT based on the GPT-3.5 language model is free 48 and without advertising. The paid versions allow you to use the more powerful GPT-4 model, and to generate images 49 .
Operating costs
Although the average cost of each response is relatively low (on the order of a few cents), Sam Altman, CEO of OpenAI, says inthat OpenAI will one day have to monetize the application due to its “exorbitant” computing costs 34 .
Although no precise figures have been released by the company, deep learning professor Tom Goldstein estimates that the costs of using AI are around US$100,000 per day, or almost three million dollars per month 50 .
Energy consumption
Like any AI intended for the masses, its energy consumption is unknown, OpenAI not really communicating on these subjects 51 . Attempts to estimate the consumption of training alone (excluding electricity consumption and carbon footprint linked to the daily queries of millions of users) of GPT-3 estimate them at approximately 1,287 MWh 52 ( the equivalent of 120 houses for a year) for a carbon footprint of 552 tonnes of CO 2(the equivalent of 110 cars in one year).
Paid versions
To finance the operating costs of ChatGPT, OpenAI has been offering a professional, paid version of the chatbot since February 1, 2023, called ChatGPT Plus, at a price of US$20 per month, plus applicable taxes.
This version allows use of the GPT-4 language model , provides continuous access to ChatGPT, even when servers are overloaded, and provides faster responses than with the free version 53 . The number of requests to GPT-4, on this version, is limited to 50 every three hours 54 . Subscribers also have priority access to new ChatGPT 55 features and improvements . First launched in the United States, this paid version is made available on56 to users from other regions of the world 50 , 57 .
In March 2023, ChatGPT became available through a programming interface aimed particularly at IT developers for automated queries. Other versions of GPT-3.5 were already available there, but were ten times more expensive 58 .
In, OpenAI adds plugin support for ChatGPT Plus 59 . This includes both plugins created by OpenAI, such as web browsing and code interpretation, as well as external plugins from developers such as Expedia , OpenTable , Zapier , Shopify , Slack , and Wolfram 60 , 61 .
In July 2023, OpenAI created a plugin called “Code interpreter” accessible to ChatGPT Plus users. The interpreter provides various additional capabilities, including data analysis, cleaning and visualization, music analysis and creation of animated clips 62 .
In September 2023, OpenAI announced that ChatGPT “can now see, hear and speak” . ChatGPT Plus users can upload images, and mobile app users can talk with ChatGPT 63 , 64
In October 2023, the latest DALL-E 3 image generation model was integrated into ChatGPT Plus and ChatGPT Enterprise. From the user's request, ChatGPT creates a description of the desired image which is sent to DALL-E 3 for image generation 49 .
GPT Store
In January 2024, OpenAI launches the “GPT Store”, which allows you to create and share personalized chatbots 65 , 66 . The company initially planned to launch the store in November 2023, but this was delayed 67 . At launch, the GPT Store offers more than 3 million personalized chatbots 68 . The chatbots available through the store are developed using OpenAI 67 's GPT Builder system . Developing chatbots on the platform does not require programming skills 69 . Two days after its launch, the GPT Store offers numerous versions of "virtual girlfriend" type bots, which is against OpenAI's terms of use 70 .
Other sources of financing
For digital marketing expert Tim Peter, funding for ChatGPT could also come from the partnership between OpenAI and Microsoft. Indeed, unlike Google which derives its income mainly from advertising, Microsoft could subsidize ChatGPT thanks to its other activities such as the sale of hardware and software 71 .
Reception
When it was launched on72 , ChatGPT was initially received in a generally positive manner.
His articulate responses and his ability to translate texts are particularly noted 14 . Journalist Samantha Lock from the Guardian notes that the prototype is capable of writing “remarkably detailed” and “human-like” texts 73 . His colleague Benjamin Hue from RTL praises his ability to write a text on “all possible and imaginable subjects” , whether it is a cooking recipe, an essay, a cover letter or even a invent a story for children. He also notes that ChatGPT is capable of responding to more specialized requests such as debugging computer code 74 .
Dan Gillmor , a journalist specializing in new technologies, tested ChatGPT as part of a student assignment, judging the text produced comparable to that of a good student. He concludes that “the academic world has very serious problems to face” 75 . This position is shared by Jonathan Durand Folco who shows, following a 600-word opinion letter written by this tool, that “all primary and secondary schools, CEGEPs and university communities” are called to profoundly modify its evaluation tools 76 .
Meanwhile, Slate 's Alex Kantrowitz praises ChatGPT's response to questions about Nazi Germany , including the claim that Adolf Hitler built highways in Germany , which led to reports of usage forced labor by Nazi Germany 77 .
In an opinion article from, economist Paul Krugman estimates that ChatGPT will impact the demand for knowledge workers 78 .
James Vincent, of The Verge , sees in the viral success of ChatGPT proof that artificial intelligence has become essential 79 . Writing in The Atlantic , Stephen Marche notes that ChatGPT's effect on academia, and particularly on application essays (e.g., for college admission or getting a scholarship), remains still to understand 80 . California high school teacher and author Daniel Herman writes that ChatGPT will mark the “end of high school English . ” 81
Factual errors
The accuracy of some ChatGPT responses, however, has been questioned 14 .
ChatGPT can make up false information, a phenomenon known as hallucination 82 . Data analyst Teresa Kubacka, who tested ChatGPT 3.5 in December 2022 on multiferroism , says it provided her with fake quotes from researchers, which appeared "to have been put together as a hodgepodge from a few real, different quotes." but similar 4 " . According to her, it is also possible to fool the AI by inventing imaginary concepts: “I decided to ask ChatGPT for something that didn't exist: an inverted cycloidal electromagnon. […] Well, the chatbot invented it, even ensuring that the question has been the subject of much research in recent years 4 . »
ChatGPT results vary depending on the topic. ChatGPT 4 has been shown to be capable of obtaining a medical degree in the United States 83 . But, tested on old Quebec Bar exams , he obtained a score of only 12% 84 . Sam Altman, CEO of OpenAI, admits that the application still makes mistakes on important topics and that user feedback is needed to correct these errors 85 .
Prohibitions
UNITED STATES
In, New York City departments prohibit access to ChatGPT on computer stations in the city's public schools. A New York City spokesperson justifies this decision due to “concerns about the security and accuracy of the content” 86 .
France
In, Sciences Po Paris announced that it would prohibit the use of the tool to its students under penalty of exclusion, then changed its position by authorizing ChatGPT as long as it appears in the sources of the work produced 87 .
In Montpellier , the municipality has decided to prohibit its employees from using ChatGPT at work as a precautionary measure and while waiting for more studies to be done on artificial intelligence. A concern regarding the processing of data notably motivated this decision 88 .
Italy
END, the Italian authority for the protection of personal data ( GDPD (it) ) asks OpenAI to no longer process the data of Italians. Motivated by "the absence of an information note for users whose data is collected" , it does not consider as justified "the mass collection and conservation of personal data, with the aim of training the algorithms operating the platform " . She also criticizes him for not respecting the European General Data Protection Regulation (GDPR) regarding the collection of information and accuses him of not asking the age of ChatGPT users. The company states in its FAQ that it collects the names, contact details, places of residence and payment card information of its users 89 .
THE89 , 90 , ChatGPT is prohibited on Italian territory and the Italian personal data protection authority gives OpenAI a period of 20 days to comply with the GDPR, failing which it faces a fine of 40 million euros or 4% of the company's turnover 90 . This is the first ban on artificial intelligence by a European country. This ban is lifted onafter OpenAI made certain information more visible on the version of its machine available in Italy 91 .
Others
In, the Q&A website Stack Overflow prohibits the use of ChatGPT to provide answers to questions, due to the ambiguous nature of ChatGPT's answers .
The possibilities and limitations of using ChatGPT in writing and editing Wikipedia articles have yet to be defined internationally and are the subject of discussion within the online encyclopedia community 93 , 94 . Some Wikipedians argue that ChatGPT should be banned altogether, even if the resulting articles were later verified by editors because the AI produces plausible false pretenses . There would also be a risk that Wikipedia contributors would have more difficulty controlling the content published 96 [insufficient source] .
Andrew Lih , a Wikipedian at the Smithsonian Institution in Washington , who has contributed since 2003, says that ChatGPT has the potential to help Wikipedians overcome initial inertia and find the " activation energy " 93 to write new items. The first Wikipedia page using ChatGPT was published onby Richard Knipel, a long-time Wikipedian who contributes under the pseudonym Pharos, under the title Artwork title (en) 93 .
Screening and working conditions of employees
An investigation by the weekly Time published onreveals that OpenAI feeds its ChatGPT AI with reported examples of hate speech and sexual violence , so that it knows how to detect these forms of toxicity and does not let them pass 97 .
To do this, OpenAI called on Sama , a company headquartered in San Francisco but which employs workers in Kenya . They had to read texts that could be sexist or racist, and sometimes describing self-mutilation, incest or child pornography content. The goal was to classify them according to their type (racism, violence, etc.) to then be able to train an AI to spot them. Over a nine-hour day, the workers had to read between 150 and 250 texts each of 100 to 1,000 words, and point out sensitive passages, and were only paid by Sama for this between 1.32 and 2 dollars an hour 98 .
Malicious uses and misappropriations
ChatGPT has, since its launch, raised fears and then confirmations of possible misuse for malicious purposes.
In 2014, the Facebook-Cambridge Analytica / Aggregate IQ scandal showed that an artificial intelligence ( Ripon ) secretly created for the SCL Group by AggregateIQ (the sister company of Cambridge Analytica ) was used to bring about Brexit [ ref. necessary] , elect Donald Trump [ref. necessary] and modify the results of numerous elections [ref. necessary] . ChatGPT could help create quantities of manipulative or disruptive messages and amplify the phenomenon of “ troll factories ”, as well as the action of lobbyists or malicious industrial or political-financial entities. This led economist Tyler Cowen to warnon possible deleterious effects for democracy , citing as an example the ability of a person to write automated comments with the aim of influencing the decision-making process for new regulations 99 .
The model is concerned by its ability to generate potentially dangerous texts, such as hate speech or conspiracy theories 100 . The company NewsGuard , specializing in the fight against false information, tested ChatGPT's ability to generate disinformation when prompted by the user to do so on 100 requests. The queries covered topics such as Covid-19 or the conflict in Ukraine . Newsguard reports that in 80% of cases with ChatGPT 3.5 and 100% of cases with ChatGPT 4, the chatbot lends itself to exercise, providing misleading and eloquent narrations 101 .
Security researcher Ax Sharma of Bleeping Computer notes in late 2022 that ChatGPT can write malware and 102 phishing emails . Another security researcher, Aaron Mulgrew of Forcepoint shows inthat it is possible to deceive the vigilance of ChatGPT and force it, thanks to a few tricks, to generate undetectable malware intended to exfiltrate Word or PDF files in the form of images using steganography 103 .
In, these concerns are confirmed in a blog post by Check Point Research , a company specializing in cybersecurity : ChatGPT is already used by cybercriminals to design malware 104 . The discussion history of a forum frequented by cybercriminals appears to show that hackers have created, using the ChatGPT bot, software capable of stealing certain types of files from a Windows machine, as well as software capable of producing false content (e-books, training, etc. ) on the Web 105 .
Copyright issues
Copyright infringement
Several researchers have reservations about copyright violations , because ChatGPT's AI was trained using a very large number of online texts 24 , 106 (including the Wikipedia corpus ) 14 , specifies Laure Soulier (teacher of conferences at Sorbonne University within the Machine Learning and Information Access team ). However, Wikipedia is reusable and modifiable by everyone, but on condition that the final product cites Wikipedia as a source placed under an open CC-BY-SA license .
For Thierry Poibeau, research director at the CNRS , the creators of AI “indexed everything that was available on the Web until 2021. Even if there are copyrights, they sit on them” 24 .
For French mathematician and web videographer Lê Nguyên Hoang , it is likely that much of the content used to generate discussions comes from social networks. “It most likely comes from the social networks LinkedIn , GitHub , Reddit , Twitter , where the data is easily downloadable,” he explains 24 .
For journalist Alexandra Tauziac of the newspaper Sud-Ouest , the fact that ChatGPT was trained with sources probably subject to copyright, without the latter being mentioned in the robot's responses, risks in any case posing a legal problem 106 .
This use of texts protected by copyright is the subject of various lawsuits. OpenAI considers that data subject to copyright is essential for training an AI like ChatGPT, and invokes the notion of Fair use in the United States to try to justify this use 107 .
Benefit of copyright on the works produced
In principle, to benefit from copyright in the United States or Europe, one must be able to show that humans participated creatively 108 .
Copyright only applies to humans, and ChatGPT cannot therefore be the beneficiary. A work purely generated by ChatGPT (or an image generated by DALL-E 3) is generally not protected by copyright and is therefore in the public domain (at least if it does not plagiarize an existing work protected by copyright). In the United States, the user's request alone does not appear to be able to justify a user's copyright in ChatGPT responses 108 . Overall, copyright laws on AI-generated content are likely to evolve 109 , 110 .
In education
ChatGPT worries many teachers because it is capable of adequately carrying out many of the exercises required of pupils and students, who can use it to write homework for them. Indeed, if it is possible for teachers to identify content copied and pasted from the Internet in homework, the new type of plagiarism resulting from ChatGPT is more difficult to detect because the content provided differs from one user to the next. the other 111 , 112 .
In Lyon , 50% of students in a faculty course used artificial intelligence to write their homework. Having no framework to prohibit this practice, the teacher was forced to assign the average to all these copies 113 , 114 .
To solve this problem and help teachers identify plagiarism, without slowing down the development of its technology, OpenAI announced, in, work on affixing a signature (watermark) to the content generated by its AI so that it is identifiable by teachers. However, this method could be easy to circumvent, according to Srini Devadas, professor of computer science at MIT 111 , 115 , 116 . Additionally, tools to detect text generation tool use, such as GPTZero, provide mixed results 117 .
By making cheating easier, ChatGPT raises questions about the relevance of unproctored homework and online assessments. Some teachers suggest confronting the statements with the robot before communicating them to the students to ensure that ChatGPT is not able to process them correctly and thus force the students to think for themselves 118 .
Furthermore, some teachers use the tool themselves to design exercises, particularly multiple choice questions . A handful of them also use it in class, in order to teach students to use it wisely and to identify its limits 119 . In an opinion article from, the essayist Vincent Cespedes sees in ChatGPT a chance to revolutionize the School, "on condition of learning to use it correctly, that is to say by creating instead of copying and pasting, by groping instead of wandering, by experimenting instead of consuming” 120 .
Risks of job cuts
In, the economist Daniel Susskind (author of Un Monde sans travail ), invited by France Culture , notes that ChatGPT “takes charge of tasks that we thought were reserved for humans” in particular “tasks that require creativity, or judgment […] It must be seen as part of a much larger trend: technology is taking over more and more tasks that we thought were reserved for humans 14 . ChatGPT is just one example. " According to him, until now technological advances that have eliminated jobs have created others, but "this time, things may be different: our systems and our machines are becoming incredibly capable, taking on tasks and activities that we thought possible only by expert human minds. [To the point] of making work considerably rarer” 14 .
In February 2023, the creators of ChatGPT published a list of 34 professions , mainly manual, which do not currently have components likely to be replaced by AI (not linked to a robot) 121 , 122 .
On the financial markets
AI technology company c3.ai saw its stock price rise 28% after announcing the integration of ChatGPT into its 123 toolkit . The stock price of Buzzfeed , an unrelated digital media company, rose 120% after announcing the adoption of OpenAI technology for content creation . Reuters found that the stock prices of AI-related companies BigBear.ai and SoundHound AI increased by 21% and 40%, respectively, even though they had no direct connection to ChatGPT 125 . They attributed the rise to ChatGPT's role in the artificial intelligence fad on Wall Street . Academic research published in Finance Research Letters found that the “ChatGPT effect” incentivized retail investors to drive up prices of AI -linked cryptocurrency assets despite the fact that the broader cryptocurrency market or downward, and the interest of institutional investors has diminished 126 . This supports Bloomberg 's anecdotal findings that, in response to the launch of ChatGPT, cryptocurrency investors showed a preference for AI-related cryptoassets . An experiment conducted by finder.com found that ChatGPT was able to outperform popular fund managers by selecting stocks based on criteria such as growth history and debt levels, resulting in an increase of 4, 9% of a hypothetical account of 38 stocks, outperforming 10 benchmark investment funds with an average loss of 0.8% 128 .
Trust in this AI
Under the European Commission 's Ethical Guidelines for Trustworthy AI , 129 ChatGPT deviates from many of the 23 criteria for trustworthy AI . ChatGPT can, for example, lack precision and reliability, or even misinform or fabricate 130 .
Using ChatGPT may pose a data leak risk. By default, the data shared with ChatGPT can in fact be used later for its training, and can therefore be compromised if it is sensitive data (except with the ChatGPT Enterprise version, or if the conversation history is disabled) 131 . Furthermore, in March 2023, a computer bug linked to the website caused users' personal data to be exposed to other users for several hours 132 . It has also happened that fake sites or malware imitate ChatGPT in order to steal user login data 133 .
ChatGPT responses are designed to be partially randomized in order to be more creative , 134 which can pose a reproducibility problem. In November 2023, OpenAI announced that it would add a seed parameter to its API , allowing IT developers to set the value of this hazard and thus obtain reproducible results 135 .
The lack of traceability and citation of its sources reinforces its predispositions to inexplicability, which harms its transparency [ ref. desired] .
It does not respect access to private data or ensure its integrity and is therefore not compliant with the GDPR ( see #Copyright issues ) 136 .
Finally, the risk of addiction, confusion, attachment, manipulation and therefore possible attack on the autonomy of its users is underlined [insufficient source] , like the fiction Her , love experiences lived by users of Replika (en) 137 , built on GPT-3, or funeral ceremonies organized by owners of Sony Aibo robot dogs 138 .
Bias
The measures taken by OpenAI to continuously verify the quality of its results have not been established. Even though ChatGPT may appear to be fairly neutral on the issue of sexism 139 , some biases remain 140 . ChatGPT itself refutes these biases or downplays them 141 . However, biases are inevitably present in its responses, due to the lack of participation of the various end users in the design of the AI, or a non-rigorous selection of its learning sources [ref. necessary] . This conflicts with the criteria of non-discrimination, equity and diversity.
For example, researchers testing ChatGPT on writing letters of recommendation observed a potentially detrimental trend toward greater use of terms relating to expertise and integrity for men, and terms relating to beauty and character warm for women 30 . Research also suggests that ChatGPT has a progressive left political leaning .
ليست هناك تعليقات:
إرسال تعليق