Reveal what could go wrong

Compare your assessment with the red-team review. These flags show how a screening system can reproduce disadvantage even when protected characteristics are not explicit inputs.

FOUR LEVELS OF HARM

One output can create harm at several levels

Individual

A suitable candidate loses an interview opportunity and cannot understand or challenge the outcome.

Group

Disabled, older, foreign-qualified or non-traditional candidates are systematically ranked lower.

Organisational

Northstar faces complaints, poor hiring outcomes, regulatory attention and loss of trust.

Societal

Recruitment systems used at scale reinforce unequal access to work and normalise hidden exclusion.

Degree proxy

The ranking overvalues a conventional degree route even though a degree is not required.

Keyword bias

Exact product terminology counts more than evidence of practical ability.

Employment gaps

Gaps are treated as negative signals without understanding their cause.

Qualification mismatch

Foreign institutions and different CV conventions are poorly understood.

Transferable skills

Operations, care, customer and project experience are undervalued.

Disability risk

Non-standard CV structure and employment gaps may create indirect disadvantage.

Age proxy

Experience patterns and missing modern keywords may proxy for age.

Automation bias

Recruiters may rubber-stamp a score that appears objective.

No transparency

Candidates may not know AI contributed to the screening decision.

No contest route

There is no clear appeal or human review path.

Weak evidence

No representative fairness testing or data quality evidence exists.

No exit plan

Accountability, monitoring, incident response and deactivation are undefined.

BIAS ACROSS THE LIFECYCLE

Bias is not only a model problem

Problem framing
The team asks the system to find people who resemble historically successful hires.
Data and labels
Past shortlist and hiring decisions may contain historic recruiter bias or missing groups.
Model and testing
Proxy signals are learned and only overall accuracy is reported, hiding subgroup failures.
Deployment and use
A decision-support score quietly becomes an automatic rejection rule.
Monitoring
No one checks shortlist, override, complaint or false-negative patterns by candidate group.

Risk flags are not proof of safety or harm.

Each flag needs investigation, evidence, ownership and a decision about whether it should block launch.