Updated by Maarten Truyens
Writing good instructions ("prompts") will make a significant difference in the results you get back from GPT3. While the entire world is still learning how to write better prompts, in many different contexts, we can already provide you some hints:
- ClauseBuddy's AI Bot will prefix the word "Draft " to your question. So if you enter "a confidentiality clause" in the box within ClauseBuddy, then GPT3 will actually receive "Draft a confidentiality clause", and start from this instruction. While GPT3 will continue to function reasonably well when the instruction contains grammatical errors, do try to avoid them.
- The best results are typically obtained when drafting in English, and translating good results afterwards using our machine translation service (DeepL). The reason is that large language models such as GPT3 thrive with volumes of data, and most of the data it is trained with, consists of English.
- Take into account that the training of GPT3 stopped several months ago, so that recent facts, legislation and case law will not be known by it...
- ... although you should be vary cautious, in general, not to rely on GPT3 for such knowledge, as the artificial intelligence is known for inventing facts ("hallucinating") and mixing up knowledge from time to time, while at the same time sounding very convincing in the texts it write.
When to use AI Bot / GPT3
Our user feedback indicates that the quality results you will get back, tend to vary between legal domains.
Due to the way GPT3 works (during its training, it essentially plays games with itself to gradually become better at guessing pieces of text), GPT3 works best when millions of paragraphs exist about a certain topic. Due to the sheer number of paragraphs available, computers can then find interesting correlations between bits of information and become very good at reproducing them. The software does not understand what it is telling you, but due to the fact that its results are a good mix of what millions of paragraphs are describing, the end result often resembles intelligence.
Applied to the legal domain, this means:
- Always, always, always check the results, and never rely blindly at them. GPT3 is extremely good at sounding convincing, even when it invents facts.
- You may want to think twice before asking GPT3 to draft advice about legal advice. Compared to, for example, domains such as advertising text, product descriptions and political discussions, both the amount and quality of legal discussions online is relatively limited.
- The warning applies even more if you are located outside the US / UK, if your legal domain is very narrow, or if your legal domain is subject to quickly evolving rules.
- For example, GPT3 can produce pretty good general disclaimers and general boiler plate clauses in English because there exist millions of examples thereof. However, don't even bother asking it for legal advice on a highly compliance-driven, fast-changing subject matter such as employment law in a small jurisdiction in the EU, because it will have limited information in the first place, and probably mix that information with employment law rules from other jurisdictions.
- For an in-depth explanation of how to write better prompts, read Microsoft's own guide.