Aipedia: Actuaries

Bias and fairness testing for AI models

Aipedia, by the Apex broking team · Last reviewed 2026-10-27

Actuaries should test AI and machine learning models for material bias and discriminatory outcomes, as TAS 100 P5.3 and FRC guidance expect. Insurer clients face Consumer Duty and equality law scrutiny, and they will look to their actuaries to show a model's outcomes are fair.

Part of: AI in Actuaries

In short

Bias can enter an actuarial model through the data, the features and the algorithm. Opaque machine learning models make it harder to see. TAS 100 P5.3 covers material bias, the FRC Models guidance calls algorithm and data bias especially significant for opaque models and expects testing by sub-group, and the IFoA guide asks members to identify, communicate and mitigate bias. Insurer clients are subject to equality law and the FCA's Consumer Duty. A biased model can lead to regulatory action and complaints against the client, reputational harm, and claims against the actuary or consultancy, which PI may respond to subject to the policy terms.

Where bias comes from

In actuarial work, bias usually arrives quietly. Historic data reflects past decisions, such as which claims were investigated or which customers were offered which products. Features such as postcode, occupation or purchase channel can act as proxies for protected characteristics. Algorithms can amplify small differences in thin segments. And the target being predicted may itself be shaped by past practice.

Traditional rating has always differentiated between risks; that is the job. The question is whether a difference is justified by genuine risk, permitted by law and acceptable to the client and its regulator. Machine learning makes that harder to answer because the reasons for a prediction are less visible. Explainability tools help, but they are only a starting point; see explainable AI in insurance.

The duties that apply

TAS 100 P5.3: material bias in models.
The FRC Models guidance: algorithm and data bias are "especially significant" for opaque models; outputs should be tested by sub-group (5.18 to 5.20). Self-updating models need drift and bias documented (4.16).
The IFoA data science and AI guide: identify, communicate and mitigate bias (2.17).
The RSS and IFoA ethical data science guide (2019) includes avoiding harm and maintaining accountability and oversight when delegating decisions to automated systems.
Insurer and pension clients are subject to the Equality Act 2010, and FCA-regulated insurers to the Consumer Duty, so they will expect advisers to support fair outcomes.

The IFoA's partnership with CISI gives members access to a certificate in ethical AI, one route to building competence in this area.

How to test

Compare predictions, prices and decisions across sub-groups, where lawful data on characteristics is available.
Where it is not, look for proxies by testing correlation between features and characteristics at an area or portfolio level.
Check model performance, not just averages, by sub-group: a model can be accurate overall and poor for a minority.
Record trade-offs between predictive power and fairness, and who decided them.
Repeat tests after retraining and monitor live outcomes.
Report the results to the client in plain terms, including what was not tested and why.

Testing is not a one-off sign-off exercise. Data, customer mix and model behaviour change over time, so the client needs a monitoring plan it can run after the actuary's engagement ends, with clear triggers for further investigation.

What can go wrong

Illustrative scenario: a consultancy builds a machine learning model for a motor insurer's new business pricing. It performs well and the insurer adopts it. A later internal conduct review finds that customers in some areas, which correlate with ethnicity, are being quoted materially higher prices than the risk data justifies, driven by an interaction between postcode and a channel variable. The insurer has to review prices and remediate customers, and asks the consultancy why sub-group testing did not pick it up.

See machine learning in pricing models and AI experience and claims analytics, where similar issues arise with fraud models.

How PI usually responds

Discrimination complaints and regulatory action are usually directed at the insurer, as the provider. The insurer may then claim against the actuary or consultancy for breach of contract or negligence, and such a claim would usually fall within actuarial PI, subject to the policy terms. Fines imposed on the insurer, and the cost of reputational damage, are generally not something the adviser's PI would pay for directly. For more on how discrimination claims involving AI are insured, see AI bias and discrimination claims; for the regulatory backdrop, Consumer Duty and PI. See also PI for actuaries.

How Apex helps

Apex arranges PI for pricing and consulting actuaries; a broker can check how your cover deals with claims linked to client conduct reviews. See the AI in actuarial work overview.

Sources

Frequently asked

Does TAS 100 require bias testing of AI models?

TAS 100 P5.3 covers material bias in models, and the FRC Models guidance says that for opaque models algorithm and data bias are especially significant and outputs should be tested by sub-group. The guidance is persuasive rather than mandatory, but it shows how the FRC expects P5.3 to be met.

How can I test for bias without data on protected characteristics?

Look for proxies. Test whether features such as postcode or occupation correlate with characteristics at area or portfolio level, compare outcomes across those groups, and check model performance by segment. Record what you tested, what you could not test and why.

Who is liable if a biased model harms customers?

Complaints and regulatory action usually target the insurer as provider. The insurer may then claim against its actuary or consultancy if the model work fell below a competent standard. That claim would usually fall within actuarial PI, subject to the policy terms.

Check your PI covers how you use AI

Tell us about your firm and the work you do. A broker will review how your current PI treats AI-assisted work, or arrange new cover. No fee for a quote. Or call 0117 325 0027.

Start a PI quote Request a callback

Apex Insurance Brokers Limited is authorised and regulated by the Financial Conduct Authority. Registered in England and Wales, company number 07014570. This page is general information about how AI is used in the profession and how insurance may respond. It is not legal, regulatory or insurance advice on your circumstances, and it does not guarantee that cover will be available or on what terms.