The short answer. Most platforms that generate care records require a staff member to sign off, which may well satisfy Regulation 36(1)(c), since by signing they adopt the entry as their own. The question is not really whether it is compliant. It is what that worker is taking personal responsibility for, and whether they read it closely enough to defend it two years later.
Last reviewed 13 August 2026 against the Children’s Homes (England) Regulations 2015. Applies to children’s homes in England.
We build AI into our own platform, so this is not a case of a supplier warning you off a capability it does not have. It is the opposite. The more we have worked on it, the clearer the boundary has become, and the more uncomfortable we are with how AI note-taking is currently being sold into care.
The pitch is always the same: staff spend too long writing up, so let the software write it for them. The time saving is real. The problem is what you are left holding afterwards.
The requirement that settles most of it
Regulation 36(1) sets three conditions for a child’s case record. It must contain the information in Schedule 3, it must be kept up to date, and it must be signed and dated by the author of each entry.
That third limb is not a formality. It exists so that months or years later, anyone reading the record can identify the person who wrote it and ask them what they meant. It is the mechanism that makes a care record evidence rather than just text.
“But the staff member still signs it off”
This is the honest counter-argument, and it deserves a proper answer rather than being waved away. Platforms that generate entries do not save them silently. A worker reviews the draft, edits it if needed, and signs. On that basis it is arguable they have adopted the entry as their own and become its author, which would satisfy Regulation 36(1)(c). We are not aware of that being tested by Ofsted or a court either way, and anyone telling you it is definitively unlawful is going further than the evidence supports.
So the useful question is not whether sign-off makes it compliant. It is what the worker is taking on when they sign, and whether the arrangement holds up under pressure. Three things follow.
Approving is not the same as writing. Reading a fluent, plausible draft and reading it critically are different activities, and the first is much easier than the second. When a draft is usually right, the checking quietly becomes a formality. That is not a criticism of care staff, it is what happens to any review step where errors are rare and the reviewer is busy. The occasional wrong entry is precisely the one most likely to pass.
The signature carries personal accountability. When a record is examined in a serious incident review or a disciplinary process, it is the person who signed it who is asked what they meant and why they wrote it that way. If the answer is that the wording came from the system, they are exposed for text they did not compose. Worth asking whether your staff understand that is what they are signing up to.
Editing pulls towards the draft. Given a generated account to correct, people fix what is wrong and leave what merely is not quite how they would have put it. The record ends up closer to the system’s version of the shift than the worker’s. Over months that shifts what the file says about a child in a way nobody chose.
None of this makes generated records unusable. It means the sign-off step is doing far more work than the sales pitch suggests, and it is worth deciding deliberately whether you want that weight resting on a support worker at the end of a difficult shift.
Where generated records go wrong
Four patterns, in rough order of how often we see them raised:
- Plausible detail that was never observed. Generated text fills gaps with what usually follows. In a care record, a sentence that reads well but describes something that did not happen is not a small error. It is a false statement about a child.
- Flattening. The specific becomes the generic. “Presented as dysregulated” replaces a description of what the child actually did, said and needed. The record survives, the information does not.
- Loss of the source. Good recording distinguishes what the worker saw from what a teacher reported from what the child said. Generated prose tends to merge all three into one confident voice.
- Uniformity that hides the outlier. When every entry reads the same, the one that should have prompted a question no longer stands out. Homogeneous records are harder to audit, not easier.
A worked example: the restraint record
Regulation 35(3) is the clearest case, because it specifies exactly what the record must contain and who must do what.
Within 24 hours the record must include eight things, among them the methods used or steps taken to try to avoid needing the measure, how effective it was, and what followed. Within 48 hours the registered person or an authorised person must have spoken to the member of staff who used the measure and signed the record to confirm it is accurate. Within 5 days they must add confirmation that they have spoken to the child.
Almost none of that is available to a model. What a worker tried before the hold lives in their head, not in any system. Whether the measure was effective is a judgement. The 48 hour step is a conversation between two people, and the sign off attests to something the signer has personally checked. A generated draft can hold the shape of this record. It cannot supply the contents.
What happens when the record is tested
Care records are written on ordinary days and read on the worst ones. A record may end up in front of a LADO, a police investigation, a serious incident review, a placing authority challenging a decision, or a court.
In each of those settings the first question about any entry is who wrote it and what they knew at the time. “The system drafted it and I approved it” is a poor answer, and it undermines every other entry that worker made, not only the one being examined.
The same applies to the independent person under Regulation 44 and to the six monthly quality of care review under Regulation 45. Both rest on records being an honest account of what happened. Analysis built on generated text is analysis of the software’s output, not of the home.
The child will read it
Regulation 14(2)(f) requires staff to help each child access and contribute to the records kept about them, and case records are kept for 75 years from date of birth. Many care leavers request their files as adults, sometimes decades later, and for some people that file is the only detailed account of years of their childhood.
There is a real difference between reading something a person who knew you wrote about you, and reading something a system generated about you. It is worth holding that in mind when weighing up twenty minutes of admin time a shift.
The data protection questions
Records about children in care are special category and highly sensitive. Before any AI feature touches them, it is worth establishing:
- Where the processing happens, and whether data leaves the UK.
- Whether your content is used to train the supplier’s models or anyone else’s, and whether you can refuse.
- Which sub-processors are involved, since the AI layer is often a third party the supplier has contracted separately.
- How the accuracy principle in UK GDPR is met when the output is probabilistic.
- Whether the feature triggers a Data Protection Impact Assessment, which processing of children’s data using new technology frequently does.
These are questions for your DPO rather than for a marketing page, and a supplier who cannot answer them in writing has told you something useful.
Where AI genuinely helps
None of this is an argument against AI in care software. It is an argument about which side of the line each use sits on. The distinction is simple: AI should help a person interrogate the record, not produce it.
Uses that sit comfortably on the right side:
- Retrieval. Finding every entry mentioning a particular concern across six months, in seconds rather than an afternoon.
- Pattern surfacing. Showing that missing episodes cluster on particular days, which a manager then investigates and explains. The insight prompts the question; the human answers it.
- Completeness checks. Flagging a restraint record with no 48 hour sign off, or a week with no entry for a child. This is the highest value use in practice and the least discussed.
- Preparation. Assembling the material for a Regulation 45 review so the registered person spends their time on analysis rather than gathering.
- Policy questions. Answering “what does our missing child procedure say about the first hour” at 2am, from your own documents.
- Dictation. Speech to text is not generation. The words originate with the worker rather than being composed for them, which is what keeps authorship where it belongs, and it genuinely helps staff who find typing slow or who write in a second language. No system can make anyone re-read what they dictated, so accuracy still rests on the worker checking their entry. The difference is that they are checking their own account rather than approving someone else’s version of it.
What to ask a supplier about AI
- Does any feature write into a child’s record without a person composing the content? If so, who is recorded as the author?
- Can that feature be turned off per home, and who controls that setting?
- Is generated text visibly marked as generated in the record and the audit trail?
- What happens if a member of staff approves something inaccurate? Where does the audit trail show that?
- Is our data used for training, and can we see that commitment in the contract rather than the marketing?
- Which model provider sits behind the feature, and where does processing take place?
Where we stand
OVcare uses AI to give managers and staff insight from their own records, and to show the reasoning behind what it surfaces so it can be checked. We do not think it should be writing the daily log or the incident record in place of the person who was on shift.
One distinction is worth stating plainly, because it is easily blurred. Assembling a report is not the same as generating a record. Our reporting, including Regulation 45 and medication reporting, collates entries your staff have already written and presents them together. There is no AI writing in that process. The words in the output are the words your team authored, which is why every figure can be traced back to the entry it came from.
That position costs us a feature we could otherwise advertise, and we would rather be straight about the trade-off than pretend there is not one. Suppliers who do generate entries are not doing something forbidden. They are moving the burden to the sign-off step. If you are considering one, the questions above are the ones to put to them.
See how the insight side works.Book a free demo and we will show you what the AI surfaces, what it does not touch, and how the audit trail reads. Our guide to record keeping requirements covers the underlying duties in full.
Frequently asked questions
Can AI write care records in a children’s home?
The Regulations do not mention AI. They require every entry in a case record to be signed and dated by its author, under Regulation 36(1)(c). Where a staff member reviews and signs a generated entry, it is arguable they have become its author, and we are not aware of that being tested. The practical question is whether the person signing has read it closely enough to defend every line of it later.
Does Ofsted have a position on AI in care records?
Inspection is concerned with whether records are accurate, attributable, up to date and reflect what happened for a child, not with the tool used to produce them. A generated record that cannot be attributed to a person will struggle on the first of those tests regardless of how it was made.
Is voice-to-text the same thing?
No, and the difference matters. Dictation captures the worker’s own account in their own words. Generation composes an account from data. Provided the worker reads and corrects the transcript before saving, dictation leaves authorship exactly where it should be.
Do we need a DPIA before using AI features?
Very likely. Processing children’s special category data using new technology is the kind of high risk processing that triggers the requirement, and your supplier should be able to support the assessment. Take your DPO’s advice rather than relying on a supplier’s assurance.
If we say no to AI-written records, how do we save staff time?
Most of the admin burden is not writing, it is duplication, hunting for documents, and reassembling evidence for reviews and inspections. Those are structural problems a single well designed system fixes without anyone signing text they did not write.
A note on scope. This article is general information about the position in England and is not legal advice or data protection advice. It does not replace the professional judgement of the registered person or the advice of your data protection officer. Requirements change, so check primary sources before relying on any point here. OVcare is a software supplier and this article is published by us, which you should weigh accordingly.
Sources: The Children’s Homes (England) Regulations 2015 (SI 2015/541), Regulations 14, 35, 36, 44 and 45, and Schedule 3. UK General Data Protection Regulation, Articles 5 and 35.
