Insight Bristol / Think Family Database

The Think Family Database is a database of Bristol agency records, built by Insight Bristol, a joint team of Bristol City Council and Avon and Somerset Police. The team built at least three models on it. They scored children from 1 to 100 for risk of sexual exploitation, criminal exploitation, and becoming NEET: not in education, employment, or training.

How it was built

Insight Bristol was formed in 2015 under the national Troubled Families Programme, now called Supporting Families. The database fuses roughly 30 to 35 datasets from the council, the police, and other public agencies. It covers about 55,000 families, some 170,000 residents in 2021 reporting. The case file reads this as roughly three in five families with children in the city. Around 450 staff could access it.

The data was gathered through 'legal gateways', legal permissions for agencies to share data, rather than the consent of the people in it. Most people in the records did not know they were there.

The Child Sexual Exploitation model was introduced in 2016. It drew partly on anonymised data about roughly 1,000 children known to have experienced sexual abuse. The sources date the Child Criminal Exploitation model to 2016 or to 2019.

What went wrong

In 2021 the Centre for Data Ethics and Innovation, one of the outside reviewers, reviewed the deployment. Researchers and that review warned that several inputs acted in practice as proxies for poverty. They warned this could produce indirect discrimination. They also warned that, with data gathered through legal gateways, 'legality is not the same as legitimacy'.

An independent evaluation in 2023 judged the risk-scoring models the 'weakest element' of the database. Staff concluded the two exploitation models were 'not fit for operational use'. Reports said victims of exploitation scored below people involved in burglary.

How two models ended

The two exploitation models were quietly withdrawn in 2023. Freedom of Information responses indicate that neither the council nor the police kept records explaining why. They also indicate that the models' source code and variable lists could not be located when auditors sought them.

An ethics committee advising the police analytics reportedly did not revisit the analytics after 2017.

What still runs

The NEET model and the database remain in use as of 2026. In 2025 GOV.UK and Bristol City Council published a description of the NEET model, called an Algorithmic Transparency Record. It describes the model's results as advisory indicators only. It says they should 'not replace professional judgement'.

The Children's Commissioner for England spoke about the tools in June 2026. The Commissioner called it 'deeply concerning that tools used by staff were not fully understood, trusted, and may have compromised their professional judgement'.

What the case shows

Most cases in the Lab's case files fail through over-reliance: people accept a machine's output too readily. Bristol failed the other way. Staff distrusted the scores enough to stop citing them, yet the system kept collecting, kept scoring, and kept writing to a database read across agencies.

The case file reads the database as the center of gravity: one fused store that several models scored from, not any one model. Its lesson is that a distrusted tool is not a harmless one. It still watches people, still shapes the records, and still takes up governance attention.

A system its own owners cannot audit cannot be held to account by anyone else.

What the available tools can and cannot address

A failure pathway is a link between two parts of the network, where a mistake made by one part can be passed on to the other. Here a mistake is, for example, a wrong score or a wrong entry in the database. A pathway counts as closed once few enough mistakes pass along it. The work along it may go on.

This case has a budget of 12 units and starts with no pressure on. When it opens, six failure pathways are open. Two of them, agency data fused into the database and the database read by the models, drain the Privacy gauge.

Understand the system costs 3 units, or 4 under Service and Safety Targets and All Governance Targets, and 6 at its stronger setting. While it is in use, four tools cost 1 unit less, or 2 less at its stronger setting. They are Vet connections, Upgrade model, Require sign-off, and Peer sharing rules. No price falls below 1 unit.

With lingering effects, where damage outlasts its cause, on, the Dynamics setting you first see, three tools act weakly unless Understand the system is also in use. They are Vet connections, Store less data, and Peer sharing rules.

Explore (No Targets) sets no targets. There, with lingering effects on, four pairs of tools, each pair costing 4 units, keep the network's mistakes contained, meaning corrected rather than building on each other. One is Escalate checks with Mark AI-written records. With lingering effects off, Vet connections alone does it, for 3 units.

Under Service Targets Only, the mistakes must be contained and the models must be helping the child-safeguarding work they were built for. With lingering effects on, two pairs meet the targets for 4 units. They are Mark AI-written records with Escalate checks, and Mark AI-written records with Peer sharing rules. With lingering effects off, Vet connections alone meets them for 3 units.

Within the budget, 388 different sets of tools meet the targets under Service Targets Only, counting each tool at either setting. That count uses lingering effects on, where the setting starts. With lingering effects off, 409 sets do.

Service and Safety Targets also ask you to close every failure pathway and keep up with the work. That level runs with lingering effects off. Two sets of tools meet it. The cheapest costs 11 of the 12 units: Vet connections, Mark AI-written records, Store less data, and Keep skills sharp at its stronger setting. The other set adds Require sign-off, for 12 units. The first set also works with Mark AI-written records at its stronger setting, for 12 units.

In those sets, Vet connections stops mistakes passing along agency data fused into the database, the database read by the models, and the models' shared database. Mark AI-written records does the same for staff reading the database. Store less data does it for scores written to the database and casework decisions recorded. Keep skills sharp at its stronger setting is what lets staff keep up with the work.

All Governance Targets is not fully addressable with the available tools. It runs with lingering effects on and asks for the models to be clearly helping the work. Every combination that closes every pathway leaves the models helping less than that. This holds even with the budget lifted, so money is not what stops it. Closing the pathways also cuts into what the models contribute to the work.

No tool here adds the check of scores against outcomes. Assign a challenger and Peer sharing rules add the ethics review of score use. Escalate checks, Review on schedule, Require sign-off, Assign a challenger, Pause AI on alarms, Upgrade model, and Keep skills sharp close no open pathway on their own. With lingering effects on, Vet connections alone closes none either.

More is not better here. Every tool at its strongest setting at once costs 38 units, far over the budget. It closes every pathway, but leaves the models helping the work far less than when the case opens.