Capacity, consent and AI

This question comes up the moment a service starts introducing AI: what do we do about people who cannot consent?

It is usually asked as one question. It is actually three, with different answers, and separating them makes the whole thing manageable.

We are not lawyers, and this is not legal advice. It is a plain reading of the Mental Capacity Act and the ICO’s guidance, with both linked so you can check us.

First: telling people is not asking them

The most common mistake is treating transparency and consent as the same thing.

Telling people is a duty under data protection law. The ICO requires that when you plan a new use of personal data, you “update our privacy information and communicate the changes to individuals before starting any new processing”. That duty applies whatever your lawful basis is, and it does not switch off because someone lacks capacity.

Consent is one possible lawful basis among several. For most care records it is not the one you rely on.

So the question “can this person consent?” does not determine whether you have to tell them. You do. What changes is how, and who else you involve.

Second: capacity is specific to the decision

The Mental Capacity Act is explicit that capacity is assessed for a particular decision at a particular time, and it starts from a presumption.

“A person must be assumed to have capacity unless it is established that he lacks capacity.” — section 1(2)

Two consequences that matter here.

A capacity assessment for something else does not transfer. Someone may lack capacity to manage their finances and have capacity to say whether they want a voice assistant ringing them on Tuesdays. Deciding they cannot be asked, because of a different assessment, is not a shortcut. It is the thing the Act was written to prevent.

An unwise answer is still a decision. Section 1(4): “A person is not to be treated as unable to make a decision merely because he makes an unwise decision.” If someone says they do not want AI anywhere near their records and you think that is irrational, that is their decision to make.

Third: you have to help people decide

This is the part most relevant to AI, and it is a duty rather than good practice.

“A person is not to be treated as unable to make a decision unless all practicable steps to help him to do so have been taken without success.” — section 1(3)

“All practicable steps” means you have to try, in the format the person can actually use, before concluding they cannot decide.

And here AI is genuinely useful rather than a complication. Producing a plain-English explanation, an easy-read version, a shorter version, or one focused only on what changes for that person — this is exactly what these tools are good at, and it directly supports the section 1(3) duty.

Three things to get right when you do it:

1. Explain the decision, not the technology. Nobody needs to understand large language models. They need to know that a tool types up what the care worker says, a person checks it, and no machine decides anything about their care.
2. Check the simplification did not remove the choice. Simplifying drops conditionals — “you can ask us not to” is the sentence that must survive. See making letters easier to read.
3. Have a person deliver it. A document is not a conversation.

When the person cannot decide, even with support

Then it is a best interests decision under section 4, and the Act sets out what you must do.

You must not decide “merely on the basis of the person’s age or appearance, or a condition of his, or an aspect of his behaviour” — section 4(1). Assuming someone with dementia would not mind an automated call, because they have dementia, is precisely the reasoning the section rules out.

You must consider whether they may regain capacity, and when — section 4(3). If a decision can wait until someone is more able to take part, it usually should.

You must “permit and encourage the person to participate… as fully as possible” — section 4(4). Partial participation still counts.

You must consider, so far as reasonably ascertainable, “the person’s past and present wishes and feelings… the beliefs and values that would be likely to influence his decision” — section 4(6). Someone who spent their life avoiding technology has told you something relevant, even if they cannot tell you now.

And you must consult, where practicable and appropriate, anyone they named, anyone engaged in caring for them or interested in their welfare, and any attorney or court-appointed deputy — section 4(7).

Note what that list is. It is people to consult, not people who decide. A daughter’s enthusiasm for an automated check-in service is a factor under section 4(7). It is not consent on her mother’s behalf, unless she holds the relevant authority.

The less restrictive option

Section 1(6) requires you to consider “whether the purpose for which it is needed can be as effectively achieved in a way that is less restrictive of the person’s rights and freedom of action.”

This is a genuinely useful test to put to any AI proposal.

• If a tool records audio inside someone’s home, is there a way to achieve the same thing without recording?
• If automated calls are proposed, is a person ringing less restrictive, and is it available?
• Does the AI option narrow this person’s choices, or widen them?

Where the less restrictive option exists and is realistic, section 1(6) says you have to consider it. Cost and convenience are real constraints, but they are your constraints, not a reason the more restrictive option is in someone’s best interests.

Where consent actually is your lawful basis

Most care records do not rely on consent. But some AI uses do, or come close — automated wellbeing check-in calls being the clearest.

If consent is your basis and a person cannot give it, consent is not available for that person. You need a different lawful basis for the data processing and a best interests decision about the intervention itself. They are two separate questions and both need answering.

Be honest about whether consent was ever real. Someone who feels a service arrived attached to their care package has not consented freely, whatever they signed.

What AI must never do

It must never make the assessment or the determination.

The Act says “the person making the determination must consider all the relevant circumstances”. A capacity assessment and a best interests decision are judgements about a particular person by someone who knows them, and they are legally attributed to a named human being.

That is the same line drawn in the golden rules and in deciding eligibility, risk or who gets a service. It applies here with more force, not less.

A practical order

1. Assume capacity for this decision. Do not carry one across from another.
2. Explain it properly, in a format the person can use. This is where AI helps.
3. If they can decide, their answer stands, including an answer you think is unwise.
4. If they cannot, work through section 4 and record it: wishes and feelings, who you consulted, why this is less restrictive or why no less restrictive option exists.
5. Tell them anyway. The transparency duty does not depend on capacity.
6. Record the decision where the next person will find it. A decision nobody can locate gets made again, differently.

For the practical side of telling people, see telling the people you support and our letter template.