Can AI Prove Coercive Control? No — But It Can Help the Court See the Pattern

The problem in one paragraph

Coercive control rarely appears in one spectacular message.

It may emerge through repeated demands, monitoring, financial pressure, parenting interference, threats, changing rules, intimidation, denigration or seemingly ordinary communications which acquire a different meaning when viewed over time.

The difficulty is not always that the evidence does not exist. Sometimes the difficulty is that it exists in 20,000 messages and nobody can see the pattern.

Credit and context

This article was prompted by a post from Jenny Rudd following a discussion at NYU London hosted by The Fifth Room.

The event brought Jenny together with family silk Charles Hale KC of 4PB to discuss AI, coercive control, message evidence and access to justice.

Jenny created Dispute Buddy after her own experience of having to collect and organise years of text messages for Family Court proceedings.

The product is designed to turn message histories into chronological reports and can optionally apply AI-assisted behavioural analysis to identify possible patterns including coercive control, financial abuse, emotional abuse, intimidation, harassment and child-welfare concerns.

JSH Law is not affiliated with Dispute Buddy. This article is an independent analysis of the evidential question the product and event raise.

Seven things to understand first

1. AI does not prove abuse.
It can assist with pattern detection and evidence organisation. The court determines facts.

2. Pattern recognition is legally relevant.
Current PD12J expressly allows examples of alleged coercive and controlling behaviour to be presented so the judge can consider the alleged abuse holistically rather than as isolated incidents.

3. A pattern found by software is still only a proposed pattern.
The underlying messages must be checked.

4. Context can reverse meaning.
A message which looks affectionate or neutral in isolation may have a very different function within an established history.

5. AI can also overread behaviour.
Ordinary disagreement, poor communication or a badly worded message must not automatically be classified as coercive control.

6. Selection matters.
If only one person’s chosen messages are analysed, the resulting picture may be incomplete.

7. The best use of AI is evidential triage, not automated adjudication.
Use it to identify where humans should look more closely.

The NYU London discussion: a problem worth solving

The Fifth Room framed the event around a simple reality.

A single message rarely tells the whole story.

Across months or years, however, communications can reveal repeated behaviour which is difficult to explain, difficult to organise and emotionally exhausting to revisit.

Jenny Rudd knows that problem personally.

She has described spending many hours collecting and organising messages for her own Family Court proceedings and wondering which communications her lawyer actually needed.

That experience led to Dispute Buddy.

The attraction is obvious.

Imagine having:

• two years of WhatsApp messages;
• thousands of SMS texts;
• group chats;
• messages about school;
• financial discussions;
• contact arrangements;
• arguments;
• apologies;
• threats;
• and apparently ordinary conversations.

A human can read all of it.

But reading every message is expensive, slow and potentially traumatic.

Computers are extremely good at processing large volumes.

That is exactly where AI ought to be useful.

First, a necessary correction to the 87% figure

Jenny’s LinkedIn post says domestic abuse appears in 87% of Family Court cases in the UK.

The underlying research is slightly narrower than that.

The Domestic Abuse Commissioner’s Everyday Business pilot examined private-law children proceedings across three Family Court sites.

Domestic abuse was raised as an issue in:

• 87% of the closed case files reviewed; and
• 73% of the hearings observed.

In the case-file study, a file counted as involving domestic abuse where it had been raised by a party or professional at any point in the proceedings.

That does not mean:

“87% of every Family Court case in England and Wales contains judicially proven domestic abuse.”

It means domestic abuse was raised in an exceptionally high proportion of the private-law children files sampled.

The distinction matters.

But the corrected figure is hardly reassuring.

It supports the Commissioner’s central conclusion:

domestic abuse is everyday business in private children proceedings.

Why coercive control is uniquely difficult to prove from communications

Courts are naturally comfortable with events.

On 14 March, X happened.

On 19 April, Y was said.

On 3 May, Z occurred.

That is useful because evidence needs particularity.

But coercive control can become invisible if the court never moves from the events to the relationship between them.

Imagine these messages:

“Where are you?”

“Who are you with?”

“Why haven’t you answered?”

“Send me a photo.”

“You know what happens when you ignore me.”

One message may be ambiguous.

Fifty messages sent repeatedly whenever the other person leaves the house may tell a different story.

Frequency matters.

Sequence matters.

Escalation matters.

What happens after resistance matters.

Whether the same behaviour occurs around money, children or social contact matters.

The evidence may therefore lie partly in relationships between messages, not simply within the wording of one message.

PD12J now expressly tells the Family Court to look at pattern

This is where technology and current Family Court procedure potentially intersect very usefully.

Practice Direction 12J was amended in 2026.

Where fact-finding is necessary, paragraph 19 directs the court to consider how the key disputed facts should be presented.

It expressly allows a schedule or table to include:

specific events and examples of any pattern of coercive and controlling behaviour so that the judge can consider the alleged abuse holistically rather than focusing upon incidents alone.

That is important.

The court still needs particular allegations.

But it also needs enough architecture to understand the alleged pattern.

That is exactly the kind of problem computational tools may help solve.

AI may be very good at finding repeated signals. The judge still has to decide whether those signals amount to coercive control.

Why message evidence can be particularly valuable

Contemporaneous messages have advantages.

They may show:

• what was said at the time;
• frequency of communication;
• timing;
• responses;
• repeated language;
• requests and demands;
• changes following separation;
• discussion about money;
• discussion about children;
• threats or consequences;
• and whether one version of events is consistent with the contemporary record.

They are not infallible.

A message thread does not show everything which happened offline.

People communicate strategically.

Someone may behave differently in writing because they expect messages to be read later.

Messages may be missing.

Conversation may move between platforms.

Voice calls may contain the important material.

And a single screenshot can remove crucial context.

But organised properly, message histories can be powerful corroborative evidence.

The hardest problem for AI is context

Jenny has herself highlighted this problem.

A message may appear friendly while carrying an entirely different meaning to people who understand the history.

Consider:

“Love you. Hope you’re safe ❤️”

In one relationship, that is affection.

In another context, after months of unwanted following, it might communicate:

“I know where you are.”

The text is identical.

The function is completely different.

This is where automated sentiment analysis becomes dangerous.

Positive words do not necessarily mean positive conduct.

Negative words do not necessarily mean abuse.

Humour, sarcasm, cultural references, code words, prior threats and relationship history may change meaning.

That is why AI should never be asked simply:

“Which messages prove my ex is abusive?”

That prompt builds the conclusion into the search.

What AI can do exceptionally well

Used carefully, AI can perform a very useful first-pass review.

It can identify:

• dates;
• speakers;
• message frequency;
• repeated phrases;
• clusters around particular events;
• financial demands;
• references to children;
• repeated questioning;
• threat language;
• changes before and after separation;
• changes before and after court orders;
• repetition across long time periods;
• and candidate examples for human review.

It can also reduce duplication.

That is enormously important.

The objective is not to give a judge 600 messages.

It may be to identify the 12 messages which best illustrate the alleged pattern while retaining access to the wider chronology.

What AI cannot safely decide

There are several lines we should not cross.

AI should not determine:

• that coercive control has legally been proved;
• that one party is truthful and another is lying;
• that ambiguous language necessarily has an abusive meaning;
• that a child is at risk;
• that conduct meets the criminal offence of controlling or coercive behaviour;
• that particular behaviour justifies a specific Family Court order;
• or that absence of a flagged pattern means abuse did not occur.

Those are evaluative decisions.

They require law, context, evidence and human judgment.

Some require judicial determination.

Others may require specialist professional assessment.

What Dispute Buddy is actually doing

Dispute Buddy describes itself as a software tool for extracting and organising message data and optionally generating AI-assisted behavioural analysis.

Its current service can create chronological message-history PDFs and can flag candidate patterns including:

• coercive control;
• financial abuse;
• emotional abuse;
• harassment;
• intimidation;
• stalking;
• and child-welfare concerns.

Importantly, its own terms contain the caveat any responsible legal-AI product should contain.

It says its AI analysis:

• may contain errors;
• may produce false positives;
• is interpretive rather than definitive;
• is not expert psychological or clinical evidence;
• and should not be the sole basis for decisions.

That is the correct boundary.

In fact, I think the strongest use case may be slightly more modest than the product’s headline question suggests.

Do not ask the system to prove coercive control.

Ask it to help identify where the evidence may justify closer human analysis.

Multilingual evidence is an interesting development

Jenny also highlighted multilingual communication.

Dispute Buddy currently says its message-history reports can contain messages in any language and that its behavioural-analysis tools support a range of languages including English, Spanish, French, German, Portuguese, Italian, Dutch, Russian, Chinese, Japanese, Korean, Arabic, Polish, Turkish, Swedish and Czech.

That could be extremely useful in cases where communication shifts between languages.

But multilingual analysis creates an additional evidential caution.

Translation is interpretation.

Meaning can depend upon:

• dialect;
• slang;
• cultural meaning;
• sarcasm;
• terms of affection;
• threat language;
• and code-switching.

If a disputed phrase becomes important evidence, the court should not simply rely upon an AI-generated English paraphrase.

Where necessary, the original language and reliable translation should remain available.

The selection-bias problem may be the biggest evidential risk

Suppose I upload five years of messages and instruct the system:

“Find evidence that this person coercively controlled me.”

The AI now has a task.

Find supporting evidence.

That risks automated confirmation bias.

A better workflow would ask:

• What repeated behaviours are present?
• What communications contradict the proposed pattern?
• Are there reciprocal behaviours?
• Does frequency change over time?
• What happened before and after the communication?
• Are apparently concerning messages isolated or repeated?
• What alternative explanations fit the text?
• Are messages missing from relevant periods?

This is the difference between:

searching for proof

and:

testing a hypothesis.

The second is much safer.

Negative evidence matters too

Good pattern analysis should not look only for confirming examples.

Suppose a party says:

“He demanded to know where I was every time I left the house.”

A message dataset can potentially help test that.

Did the behaviour occur:

• three times?
• thirty times?
• only during an identified emergency?
• only while a child was expected home?
• before separation?
• after separation?
• after a non-molestation order?

Frequency and absence can both matter.

That is one of the reasons large-corpus analysis could become so useful.

Do not lose the source while building the analysis

This is fundamental.

An AI summary is not the underlying evidence.

A behavioural-analysis table is not the underlying evidence.

A chronology is not the underlying evidence.

The messages are.

Where an important proposition is relied upon, it should be possible to trace backwards:

Analysis → selected example → timestamp → message thread → original source.

If that chain breaks, the evidential value falls.

The original source should therefore be preserved wherever possible.

That may include:

• the original device;
• native message history;
• full surrounding conversation;
• timestamps;
• speaker attribution;
• attachments;
• and any export used to generate the report.

The aim is reproducibility.

Another person should be able to inspect the proposed pattern and test whether the source supports it.

Witness evidence still belongs to the witness

Family Procedure Rules remain relevant even when technology assists the preparation.

Under PD22A, a witness statement should, where practicable, be in the witness’s own words.

It must distinguish between matters within the witness’s own knowledge and matters of information or belief, identifying the source of the latter.

The witness verifies the statement with a statement of truth.

That obligation does not disappear because AI helped organise the evidence.

If software suggests:

“This demonstrates a sustained pattern of intimidation”

the witness should not simply paste that sentence into a statement.

The better evidence may be:

“Between January and March, I received 42 messages asking where I was or who I was with. Examples are at [references]. When I did not respond, the messages became [describe source accurately].”

Then the court can evaluate the evidence.

These datasets are exceptionally sensitive

Years of intimate communications may contain:

• children’s information;
• health information;
• addresses;
• financial information;
• sexual material;
• third-party information;
• school information;
• and allegations of criminal conduct.

So any AI workflow needs serious attention to privacy and security.

Dispute Buddy’s current privacy policy says ordinary message processing occurs locally on the user’s device by default.

Where a user requests behavioural analysis, selected message data are transmitted to its cloud infrastructure and processed using a third-party AI provider, currently Anthropic.

The company says the content is deleted after processing and is not retained for training.

Those are useful safeguards.

But users should still understand:

• what data leave their device;
• which provider processes them;
• where processing occurs;
• what files remain locally afterwards;
• and whether using a work or shared computer creates additional risk.

Domestic-abuse evidence requires security-by-design, not security as an afterthought.

From “AI found a pattern” to court-ready evidence

This is where the legal work begins.

A useful pathway might look like this:

The JSH Law AI-assisted coercive-control evidence workflow

I would structure the analysis around six questions.

1. What happened?

Identify the communication or event objectively.

2. Where is the source?

Link directly back to the original message or record.

3. Is it repeated?

How many similar examples exist and over what period?

4. What changed?

Did frequency, tone or consequence alter before or after separation, proceedings, new relationships, orders or resistance?

5. What is the alleged function?

Monitoring? Isolation? Financial restriction? Intimidation? Undermining parenting? Use of children? Something else?

6. What is the welfare consequence?

Why does the alleged pattern matter to the child’s welfare or the decision the court must make?

That produces evidence architecture rather than an AI accusation.

Thousands of messages and no idea where to start?

This is exactly the kind of evidence problem JSH Law works with.

Defined-scope support can include:

• message and email chronologies;
• evidence-source mapping;
• coercive-control pattern analysis;
• checking AI-assisted outputs against the source;
• allegation schedules;
• witness and position statement preparation support;
• Cafcass report analysis;
• safeguarding evidence organisation;
• court-bundle preparation support;
• and hearing preparation.

The objective is not to ask AI to decide whether abuse occurred. It is to make the relevant evidence easier for a human decision-maker to see and test.

Book a consultation

A practical example: “Where are you?”

Imagine the dataset contains the phrase:

“Where are you?”

AI identifies it 64 times over 18 months.

That sounds significant.

But it is not yet evidence of coercive control.

Now test it.

Of those 64 messages:

• 12 relate to collecting the child;
• 8 relate to ordinary travel plans;
• 44 occur when the recipient is socialising independently;
• 31 are followed within minutes by repeated demands for a reply;
• 17 are followed by accusations;
• 9 include a threat or consequence;
• and the frequency rises sharply after separation.

Now the pattern is more informative.

But we still ask:

• What did the recipient say?
• Was location sharing previously agreed?
• Are there messages which contradict the interpretation?
• Was there a legitimate reason for concern?
• Are both parties behaving similarly?

Only then do we have something capable of meaningful human analysis.

This illustrates the distinction perfectly.

AI has not proved coercive control.

It has made a potentially important evidential question visible.

The strongest safeguard may be asking AI to argue against itself

One of the easiest ways to misuse AI is to ask only confirmatory questions.

So once a candidate pattern has been identified, ask:

“What evidence in this dataset is inconsistent with this interpretation?”

“What benign explanations are reasonably available?”

“Which examples are ambiguous?”

“Which examples rely upon context not contained in the messages?”

“Are there periods when the alleged pattern does not occur?”

“Does the other party make similar communications?”

That does not make an AI system objective.

It does make the human workflow less vulnerable to confirmation bias.

And “gaslighting” needs particular caution

Tools commonly use categories such as “gaslighting”.

That language may be useful descriptively.

But it can also become too easy to apply.

Two people disagreeing about what happened is not automatically gaslighting.

A mistaken recollection is not automatically gaslighting.

A contradiction is not automatically gaslighting.

The evidentially useful question is more specific:

What was said, what objective evidence exists, was the same tactic repeated, and what effect is alleged?

Labels should follow evidence.

They should not replace it.

Could AI reduce the trauma of evidence preparation?

This may be one of the most valuable aspects of the technology.

Victims of domestic abuse can be required to revisit years of communications repeatedly.

Search the message.

Read it again.

Screenshot it.

Remember what happened.

Explain it to the solicitor.

Put it into a statement.

Read the response denying it.

Prepare for cross-examination about it.

Technology capable of organising messages without requiring a person to manually reread every communication could reduce some of that burden.

That is worth taking seriously.

Trauma-informed technology does not mean lowering evidential standards.

It means designing the process so that rigorous evidence preparation creates no more distress than is necessary.

So, can AI prove coercive control?

No.

And I would be uncomfortable with any legal-technology product being treated as though it could.

But that is not an argument against the technology.

It is an argument for using it properly.

Coercive control is one of the areas where AI may offer something genuinely valuable because the evidential problem often involves:

• volume;
• repetition;
• change over time;
• relationships between communications;
• and patterns which are difficult for one human to detect across thousands of records.

Those are computational strengths.

But proving the allegation requires something else.

It requires:

• authentic evidence;
• context;
• legal relevance;
• fairness to both parties;
• alternative explanations;
• human interpretation;
• and ultimately judicial judgment.

Use AI to find the possible pattern.
Use evidence to test it.
Use human judgment to understand it.
Let the court decide whether it is proved.

That is the model I think family justice should be building towards.

And Jenny Rudd is right about the underlying problem.

People should not have to spend forty hours manually trawling through painful messages simply because our evidence systems were designed before technology could do the first pass for them.

The opportunity is not to automate findings of abuse.

It is to make relevant evidence easier to find, easier to organise and easier to test.

That could improve access to justice.

And if we build it carefully, it could make the court’s job easier too.

Related JSH Law analysis

• Seeing the Pattern: What the Major Coercive-Control Frameworks Can — and Cannot — Tell the Family Court
• From Incidents to Trajectories: Coercive-Control Evidence and Behavioural Patterns
• AI Evidence in the Family Court: Deepfakes, Screenshots & Digital Evidence
• Can AI Tell the Truth? Why the Difference Matters in Law
• The JSH Law Six-Question Check

Sources and further reading

• The Fifth Room — “Can you prove coercive control with AI?”, NYU London, 6 October 2026.
• Jenny Rudd / Dispute Buddy — product information, methodology information and privacy documentation.
• 4PB — Charles Hale KC professional profile.
• Domestic Abuse Commissioner — Everyday Business: Addressing Domestic Abuse and Continuing Harm Through a Family Court Review and Reporting Mechanism.
• Family Procedure Rules — Practice Direction 12J.
• Family Procedure Rules — Part 22 and Practice Direction 22A.