Law for filmmakers and entrepreneurs

Blog

Using AI for Contracts: What It Can and Can't Do

Recently I gave a talk called Taking the Fear out of Contracts, covering the key commercial terms in a contract, the small print and some negotiation basics. I also touched on when to use or not use AI. Most of the questions afterwards were about AI.

I do use AI. I have a legal LLM that helps me with my contracts, and most of my clients use ChatGPT, Gemini or Claude for at least part of their business. Used in the right way, it is a useful tool.

The difficulty with contracts – and anything else, really - is that you need to be able to assess the answer an LLM gives you. In other words: I don't use an LLM for my accounts, because I know very little about accounting and would not be able to tell whether its answer was correct. The same applies to contracts: if you can't review it, don't trust it.

In this blog, I set out where AI can help with your contracts, where it can't, and what to check before you upload anything.

Can I use AI to draft my contract?

In my view, not yet – or at least not as a first draft. There are three main reasons.

First, an LLM will try to give you what you ask for. If you ask for a contract on two pages, it will give it to you on two pages, even where the transaction can't properly be documented in two pages. The worst example I have seen was a share purchase agreement. The sale and purchase of shares is a complex and highly regulated transaction, and my client's LLM had condensed it to one page. Typically, those agreements aren’t less than 40 (!) pages long. As a result, the shares were not validly transferred, and fixing it would have been costly, not least because some registrations had already been altered on the basis of that dubious one pager.

Second, LLMs can get the industry wrong. An option agreement in the film industry – where a producer pays for the exclusive right to acquire the rights in a book or screenplay within a set period – is a very different document from an option agreement in property law. Unless you tell the LLM which one you mean, it may not know. AI also doesn’t necessarily know the ins and outs of your industry. It might be completely customary in publishing to never waive the author’s moral rights. Yet in film and tv every financier’s or broadcaster’s chain-of-title checklist contains a waiver of moral rights.

Third, LLMs are designed to be helpful, which in practice often means agreeable. I just tried again and the business model (!) of Claude is apparently still trained to provide me with a one page “share purchase agreement”, as well as a one-page option agreement. Beyond laughable and most definitely not fit for purpose.

Where can AI help?

Where I think AI can be useful is in reviewing a contract that you have been sent. If you receive a 40-page agreement full of small print, an LLM can translate the legalese into plain English and summarise the key commercial terms – who is doing what, who is paying what, and when. Some of the summaries I have seen have been good. Reading the summary doesn’t mean that you should rely on it – you still need to read the contract yourself and think about whether what it is asking of you is as agreed. However, it can be a useful starting point.

To get the most out of it, I would use my SMARTT framework and ask one question per chat, with the chat memory switched off. That means that the LLM starts ‘afresh’ each time you ask it a question. If you stay in the same chat window, it will use the previous exchange as ‘context window’ for your next questions. That wastes energy and also increases the chances that the LLM might mix things up or start hallucinating.

Here is my SMARTT framework – use one chat per category and ask the LLM to link its answers to clause references, so you can double check its answers.

1.     Scope – what is the scope of the contract? What does it say I will do, and what does it say the other party will do?

2.     Money – who pays what, and when? What do I have to pay, and when do I get paid?

3.     Assignment of rights – are any rights being assigned or licensed, and if so, which ones? Is the licence exclusive?

4.     Risk – which risks am I taking on, and which is the other party taking on? You can also ask how a risk could be mitigated – for example, by splitting the fee 50% upfront and 50% on delivery.

5.     Timelines – what are the deadlines for each party?

6.     Termination – if things go wrong or if one party wants/needs out, what are the mechanics for terminating the contract? And what happens following termination? Where will the rights sit? Will the parties have any further obligations towards each other?

It also helps to give the LLM some context at the beginning of each chat: who you are, which party you are in the contract, which industry you work in, and which law governs the contract. I wouldn’t start off giving the LLM your personal preferences or red lines as you then risk the LLM telling you what it thinks you want to hear.

With all of the AI’s answers: read them carefully and think through what that means for your business relationship. A contract should be telling the story of what you and your business partner have set out to do, how and when. So think about the practicalities: can you do what the contract tells you to do? Can you fulfil your role? Is the contract correctly talking about what you expect the other side to do?

Lastly, check whether what the LLM is telling you is practical or doable. An LLM recently advised me to make a securities filing in Illinois on a very simple deal that happened to have a tangential link to Illinois. I will not be doing so. Crucially though – I was able to assess from a legal point of view whether the securities filing in Illinois was actually required. If in doubt: ask your lawyer, because your LLM ain’t one.

What are the hidden traps?

Confident mistakes.

I recently asked ChatGPT whether one of my standard film contracts contained any contradictions. It told me that clause 6.1 promised a credit, but clause 13.3 said that failing to give the credit was not a breach, and that this needed to be amended. It was wrong on two counts. The clause it was referring to was clause 12, not clause 13. And it wasn't a contradiction: it is standard in film and TV contracts to promise a credit and then provide that an inadvertent failure to give it isn't a breach. Once a film has been released, the credits can't easily be recalled and corrected – this was true of film reels, and it is still largely true of digital files.

LLMs recognise patterns in language; they don't know things. They never went to law school. They also never had to live in the real world – aptly demonstrated by this example:

A few months ago, a prompt circulated asking the LLM whether you should walk or drive to a car wash 100 feet away in order to wash your car. Pretty much all models replied with a version of “It's just 100 feet away, you should walk”. As this went so viral, the models have now learned that they gave the wrong answer. However, the consistency and confidence of this completely non-sensical reply should give us some food for thought. Should you trust a model that tells you to walk to the car wash if you want your car washed with a contract? (Spoiler: no. And not with any other advice that you can’t check, either).

Polished formatting.

LLM-drafted contracts are usually very well formatted – often better formatted than a lawyer's, mine included. The difficulty is that a well-formatted contract doesn't look as though it contains mistakes. It looks ready and polished. Truth is: they all look better than what’s hidden inside. A famous German author once said that he writes everything in pencil because seeing his words on screen makes them look ‘final’. I wouldn’t go as far as drafting my contracts in pencil (nor would my clients be happy to pay for that time) – but reflecting on style over substance is probably not a bad thing.

Industry standards.

As mentioned above, in English film and TV contracts, authors are normally required to waive their moral rights. That is not the case in publishing, for example, or in many other jurisdictions. If an LLM suggests striking out the moral rights waiver and you raise this with the other side, it is unlikely to be taken seriously, and it may weaken your position on the points that matter. The lawyer or experienced industry professional on the other side will spot from a mile away that you’re a bit green behind the ears.

Skimming the middle.

In my experience, LLMs read the beginning and the end of a long document more carefully than the middle. That’s based on how they were trained to summarise books or articles – the important parts are often in the introductions and conclusions. However, in a contract, the middle is often where the warranties and indemnities are. Not a small thing to be skipping! Asking targeted questions in separate chats, as above, makes it more likely that the whole contract is read.

Signed means binding.

If you sign an LLM-drafted contract and so does the other party, you generally have a binding contract. Not having read or understood it is not a defence. This cuts both ways. The contract may commit you to obligations you didn't intend to take on. Equally, it may not achieve what you wanted it to achieve – as with the one-page share purchase agreement above, where the shares were not validly transferred. In both cases, you are likely to find out only when something goes wrong.

Is it safe to upload a contract to an LLM?

Before you upload anything, check your privacy settings. Most of the main LLMs have a setting that determines whether your chats can be used to “train their models” (aka read and save your chats!), and the defaults vary between products and plans. I would only use models and plans that allow you to switch off the ‘training’ or ‘improving’ options.

There are two reasons for this. Contracts are usually drafted by someone else, and most lawyers restrict you from using the contracts that they drafted for you in different contracts. We’ve all seen and used recycled contracts, however uploading someone else’s copyright to a machine which then uses someone else’s IP to spit out a similar version to another user doesn’t seem to be in the spirit of what lawyers provide you with.

Secondly, contracts often contain personal data. A cast agreement, for example, may include a date of birth, a home address or even a passport number, and the data protection laws apply to how you handle that information (I have written about data protection for producers here).

And yes, there have been publicised incidents where parts of personal chats were inadvertently made visible to other users, so I would work on the basis that anything you upload could be seen by someone else.

At the very least, I would anonymise the contract and remove any personal data before uploading it – and only to a model which allows you to switch off the ‘training’ or ‘improving’ option.

What about scripts and manuscripts?

I would not upload them. Assuming that you may still do so - if you do, I would only do so with the ‘training’ setting firmly switched off. 

Interestingly, the industry is already adapting. A production company I spoke to recently has stopped sending scripts as attachments altogether. Instead, they send a link to a non-downloadable copy, which expires after a set timeframe.

If you do circulate scripts on email, to protect yourself a tiny bit, I would add a clause to your NDA prohibiting the recipient from uploading it to an LLM.

What about AI tools designed for lawyers?

I was also asked about AI tools built specifically for lawyers. I have tried a few. The first was expensive and, in my experience, hallucinated too often to be worth it. Sort of like a paralegal who did great work 50% of the time and was completely drunk (or stoned?) the remaining 50%. The second was more useful, mainly because it worked within Word and produced tracked changes. It was still a bit hit and miss and apart from pushing my brain into gear (useful at the end of the day), it still required me to correct about 50% of its work.

I now use the business version of Claude, which also has a Word add-in and costs considerably less. I find its proposed tracked changes particularly helpful, because I have to accept or reject each change rather than assuming the document is finished. It still makes mistakes – the Illinois filing was Claude – but a trained reader can spot and question them. My personal take is that these general tools will make many of the specialist AIs redundant. Whichever tool you use: you do need to be able to review the work it provides you with.

Next steps

The overarching point is a simple one: if you can't review what an LLM has produced, don't trust it.

If you are setting up your business and need contracts and can’t afford a lawyer for all of the legal work, I would:

1.     Start with a reliable first draft.

Use a template from a reputable provider, ask trusted colleagues in your industry what they use, or ask a lawyer to review the template you intend to use. If it will form the basis of most of your work, it is worth the investment.

2.     Use AI to help you understand and adapt it

Plain-English summaries, suggested wording and the SMARTT questions above. If you have a model that offers a Word plug-in, use it to provide you with tracked changes. Don’t replace your brain. If you can’t assess or review the output – don’t use it.

3.     Don’t rely on it for negotiation advice.

In my experience, an LLM will tell you that you are right. Before any negotiation, I would work out your must-haves and red lines, open somewhere above them, and listen to your instincts.

 

In my experience, women have a particularly hard time to start with ‘asks’ above their ‘take’ price. But basically: if you’re happy to settle for £10, then don’t start with £10. Maybe don’t start with £20 either (as that risks alienating the other side) but somewhere close-ish to where you want to land. So maybe somewhere between £12 and £15 and then trade that in for something else you want to keep as part of your red lines that you need to have established before you went into the negotiations.

Producer Contract School

I am also launching Producer Contract School, with monthly classes, a checklist for each lesson, recorded sessions and a monthly live Q&A with me. Sign up to the waitlist here.

A bit further down the line: my business contract school will be launched in 2027.

If you have used an LLM to draft or review a contract and aren't sure it does what you need it to, get in touch and let’s have a chat.

NB: I am qualified to advise on matters of English (UK) law. If you need help with matters governed by another jurisdiction, I am afraid I will not be able to assist, but I'm happy to point you in the right direction.

Silvia SchmidtComment