#EdTech AI Bias and Fairness in School Marking
Intro
By the time schools reach the point of rolling out AI properly, most of the big questions have already been discussed. Safeguarding, GDPR, acceptable use — all essential groundwork.
But there’s another issue that deserves more attention, especially when AI is used anywhere near assessment or feedback: bias.
AI tools don’t mark work in the way a teacher does. They don’t understand context, lived experience, or the subtle progress a pupil might be making. What they do is spot patterns based on the data they were trained on — and that’s where problems can creep in.
The risk of bias in AI marking and feedback
AI bias isn’t usually deliberate. It’s often the result of skewed or incomplete training data. If an AI system has mostly been trained on narrow or unrepresentative examples, pupils who don’t fit those patterns may be judged unfairly.
A practical example
In one secondary school, staff trialled an AI tool to help draft feedback on short written answers. At first, it saved time and produced consistent comments. However, teachers began to notice that pupils who used more conversational or regional language received harsher feedback than peers who wrote in a more formal academic style.
The issue wasn’t the pupils. Most of the example material used to guide the tool came from older, high‑attaining sets, with limited variation. Once teacher review was made mandatory and the tool's use was adjusted, the issue was resolved.
Clean data matters more than clever tools
AI systems are extremely sensitive to input quality. Out‑of‑date rubrics, inconsistent examples or poorly reviewed material can all distort outcomes. Good results depend on well‑governed, relevant data — not just advanced technology.
Where AI fits (and where it doesn’t)
Used carefully, AI can support feedback drafting, spot common gaps and help with consistency checks. It should not be used for final grading, high‑stakes assessment or decisions where personal context matters.
What this means in practice (DfE‑aligned)
Schools using AI in feedback or assessment should be able to demonstrate that:
Professional judgement remains central
AI is used only to support teachers (for example, drafting feedback or highlighting common gaps), not to make final grading or high‑stakes decisions.
Data protection and fairness are considered together
Training examples, prompts, and uploaded materials are relevant, up to date, and reviewed regularly to reduce the risk of bias or unintended disadvantage.
Use is proportionate and transparent
AI tools are introduced with clear boundaries, documented use cases, and an understanding of when they should not be used — particularly in summative assessment.
Staff are trained to challenge outputs
Teachers are supported in questioning AI‑generated feedback, recognising potential bias, and overriding or amending outputs where context matters.
Oversight and review are ongoing
AI use is treated as an evolving practice, with periodic review rather than a one‑off implementation, in line with wider digital governance expectations.
This approach aligns with DfE expectations around safeguarding, data protection, assessment integrity, and responsible use of digital technology.
Training staff to spot bias, not just use tools
Schools handling this well focus on staff confidence as much as capability. Teachers are encouraged to question AI outputs and treat them as starting points, not answers.
Where Tech Shepherd fits
This is where many schools pause — not because they don’t see the value of AI, but because they want to roll it out properly.
At Tech Shepherd, we support schools by setting clear boundaries around AI use, reviewing data sources before introducing tools, and helping staff feel confident using AI in ways that support good teaching without creating new risks.
Conclusion
AI can support feedback and consistency, but it should never replace professional judgement in marking
Bias often comes from training data and poor inputs, not bad intent
Schools need clean data, clear boundaries and staff confidence to challenge AI outputs
AI should support marking, not replace teacher judgement
Bias is often linked to data quality and configuration
The strongest rollouts focus on people, process and ongoing review
Comments