ChatGPT opens to the public

On 30 November 2022 the American company OpenAI made ChatGPT available to the public, a language model tuned for dialogue, as a free research preview. It belongs to the GPT-3.5 series and was refined using reinforcement learning from human feedback. According to OpenAI, it can answer follow-up questions, admit its mistakes, challenge incorrect premises and reject inappropriate requests.

Within hours, users around the world were sharing on social media examples of texts, code and summaries generated by the tool. For companies and law firms the interest is obvious: automating drafts, summaries and queries. But it was born as a trial, with limitations its own developer acknowledges and with open legal questions about data, copyright and liability.

What it is and is not

A language model learns to predict the next word from enormous amounts of text, and produces answers that sound coherent. It does not consult a database of verified facts or tell truth from falsehood by itself. That is why it can produce plausible but incorrect or nonsensical answers, a limitation OpenAI pointed out when presenting it.

Results vary with how the question is put, and its knowledge has a time limit, because it was trained on data predating its launch. It is a good tool for organising ideas or sketching a draft; it is not, on its own, for legal, tax or financial decisions, or for citing rules or judgments without checking them. Nor does it guarantee that it cites real sources: it may invent plausible references or data, so every relevant statement must be checked against reliable sources.

Data and confidentiality

Whoever types into the tool is sending information to a third party. Conversations may be reviewed by the provider to improve its systems, so it is wise not to enter personal data of customers or employees, trade secrets or information subject to a professional duty of confidentiality. The General Data Protection Regulation applies when personal data are processed, also with external tools. It should also be remembered that what is pasted into the tool may include third-party data, such as contracts or emails, over which the company has duties of confidentiality and protection.

Companies with confidentiality obligations, such as law firms, consultancies or health bodies, should decide whether to allow its use and on what terms. There are also open questions about who owns the rights to the texts generated and about the materials the model was trained on, which courts and legislators are beginning to address.

Rules and good practice

The European Union has been processing since 2021 a proposal for an artificial intelligence regulation that classifies systems by level of risk and is still under negotiation, so there is not yet a specific European rule on these tools. The proposal distinguishes prohibitions, high-risk systems with strict requirements and transparency obligations for uses such as chatbots, which should tell users they are talking to a machine. Meanwhile the general rules on data protection, intellectual property, consumer protection and liability apply.

A short internal policy may be enough to start: which uses are permitted, what information must not be entered, who reviews results before they go to a client and how it is recorded that the tool was used. It is also worth training teams and checking client contracts in case they restrict third-party use. Finally, watch the intellectual property of what is published and the company’s reputation if a generated text contains errors. This is general information.

The usefulness of these tools grows fast, but responsibility for what is signed remains human.

This article is general information. It is not legal, tax or investment advice, and it does not take anyone’s particular circumstances into account.