Teachers at Euan Blair firm report ‘horrendous stress’ after AI used to rate their work
By Gammatek ISPL| Last updated: September 2026 | 14 min read
Why This Matters to You
If your company runs, or is about to buy, enterprise performance management software, the story of teachers at a UK tech training firm is a warning shot. Staff there say an AI that scores their classes has left them on edge, sleeping badly and, in some cases, seeking therapy. The same category of tool is being sold to every large employer, and the mistakes behind this rollout are easy to repeat. Get it wrong and you risk burnout, attrition, legal exposure and a public backlash. Get it right and AI can genuinely improve coaching and quality. This article explains what happened, why the design of these systems produces stress, what regulators are signalling, and the questions to ask before you sign a contract.
What Happened at Multiverse
According to reporting by The Guardian, teachers at Multiverse, the tech training company co-founded by Euan Blair and recently valued at £1.6bn, have blown the whistle on a new AI monitoring system. The company delivers apprenticeship training in computer and AI skills to thousands of UK public and private sector workers. Transcripts of online classroom sessions are now analysed and scored by an AI, which alerts human managers when it suspects an instructor has done something wrong.
The reported triggers are strikingly granular. The system is said to flag problems as specific as failing to resolve an online connection disruption within roughly a minute, or leaning on filler phrases such as "sort of" and "kind of". Each instructor is reportedly tagged with a "risk status", given a percentage confidence score, and accompanied by written comments that managers can scan at a glance. Teachers described the system as remorseless and unnerving, and some reportedly told the newspaper of "horrendous stress".
Some context matters. The Guardian reports that the firm uses a mix of AI models, including products from Anthropic and OpenAI, and that the monitoring agent was introduced only in recent weeks. It also reports that the education inspectorate Ofsted found the company fell short of expected standards in several areas in February, while praising the instructors themselves as skilled and effective. The instruction set given to the AI reportedly tells it to judge patterns across a whole transcript rather than isolated slips, which suggests the designers anticipated over-flagging. Before you republish or build on this story, read the full Guardian piece and check for any response from the company, since coverage is still developing.
[IMAGE 2: timeline graphic of the story's key facts, minimal text, original design] Alt text: Simple timeline showing introduction of AI teacher scoring, staff reaction, and wider debate about workplace AI monitoring Caption: The rollout was recent; the reaction was fast. That gap is the lesson.
This Is a Pattern, Not an Outlier
Multiverse is not alone. The same reporting notes that earlier this year Burger King began using AI to track restaurant workers' customer interactions, and that this summer Meta paused tracking of employees' computer keystrokes after data privacy concerns and staff backlash. Three very different workplaces, one shared design idea: capture behaviour continuously, score it automatically, and route exceptions to a manager.
That idea has a name in management research: algorithmic management. It is attractive for reasons every operations leader recognises. Human observation is slow and inconsistent. A manager can sit in on perhaps one session per instructor per month, while software can review every session. Vendors of enterprise HR software and enterprise human resources software pitch exactly this: complete coverage, consistent standards, lower cost per review.
The trouble is that coverage and consistency are not the same as accuracy or fairness. And when the thing being covered is a person's daily working life, the psychological cost of constant coverage is itself a business risk.
Inside the Machine: How AI Performance Scoring Actually Works
To evaluate these tools you need a mental model of the pipeline. Most transcript-scoring systems follow the same five steps.
[IMAGE 3: five-step pipeline diagram: capture, transcribe, score against rubric, flag and rank, manager review]Alt text: Diagram of an AI performance scoring pipeline from session capture to manager dashboard Caption: Every step introduces a place where errors and stress can enter.
Capture. Sessions are recorded or live-streamed.
Transcribe. Speech becomes text. Accents, jargon and poor audio all degrade this step, and errors here flow silently downstream.
Score against a rubric. A language model is given written standards and asked whether the transcript meets them. Rubrics are converted into prompts, so vague standards become vague judgments.
Flag and rank. Outputs are turned into a status label and a confidence figure.
Manager review. A human sees the flag, often as a dashboard summary rather than the underlying evidence.
Each step has a known weakness, but three deserve special attention because they explain the stress teachers describe.
Weakness 1: Proxies masquerading as quality
Counting filler words is a proxy for "has this instructor internalised the material". The proxy is not worthless, but it is weak. Fluent experts hedge accurately when a question is genuinely uncertain, and nervous novices can sound crisp while being wrong. Once a proxy becomes a target, people change their behaviour to satisfy the metric rather than the goal, a pattern known as Goodhart's law. Teachers who start scrubbing "kind of" from their speech are optimising for the scorer, not the students.
Weakness 2: The confidence score illusion
A "percentage confidence score" looks like precision. In practice it usually reflects how strongly a model's output matches a pattern, not a calibrated probability that a real problem occurred. Managers who see "87% confidence" tend to read it as "87% likely this teacher is failing", which is a different claim entirely.
Weakness 3: Base rates and false alarms
This is where simple arithmetic changes the picture. The following is an illustrative calculation, not data from Multiverse. Suppose a flagging system correctly catches 90% of genuinely problematic sessions and wrongly flags 10% of fine ones, which sounds impressive. Suppose only 5% of sessions truly have a problem.
Out of 1,000 sessions, 50 are truly problematic and 45 of those are caught. Of the 950 fine sessions, 95 are wrongly flagged. That means 140 sessions are flagged, and only 45 of them, about 32%, are real problems. Roughly two flags in three are false alarms.
Now look at it from the teacher's side. If an instructor delivers 20 sessions a month and each carries a 10% chance of a false flag, the probability of at least one false flag in a month is about 88%. A perfectly competent teacher should expect to be flagged most months. That is a recipe for anxiety, and it explains why "unnerving" is a rational reaction rather than an oversensitive one.
[IMAGE 4: simple bar diagram showing 1,000 sessions split into true flags, false flags, missed, and clean] Alt text:Bar diagram illustrating how false alarms outnumber true flags when real problems are rare Caption: Illustrative example: even a system that sounds accurate can produce mostly false alarms when real problems are rare.
Comparing Monitoring Approaches
Not all workplace analytics carry the same risk. The table below is an original comparison based on the design features discussed above. It is a framework for discussion, not a benchmark of specific products.
Approach | What it measures | Stress risk | Legal and privacy exposure | Typical failure mode | Best used for |
Continuous transcript scoring | Speech patterns, compliance with scripted behaviours | High | High (recordings, profiling, possible high-risk classification) | Proxy gaming, false alarms, chilling effect | Narrow quality checks with human review |
Keystroke and screen tracking | Activity levels, application use | High | High (privacy backlash, disproportionality) | Measures busyness, not value | Rarely justified outside security contexts |
Outcome-based analytics | Learner results, customer satisfaction, error rates | Low to medium | Medium | Lagging signals, small samples | Performance management with context |
Sampled human observation with AI assist | Selected sessions, reviewed by a trained coach | Low | Lower | Slower, less coverage | Coaching and development |
Self-review tools | Teacher-initiated feedback on their own sessions | Low | Low | Lower adoption | Professional growth |
The pattern is clear. The tools that maximise coverage also maximise stress and risk. The tools that maximise trust tend to put the employee in control of the data.
The Legal Picture Employers Cannot Ignore
Regulators are paying close attention to exactly this use case.
European Union. The EU AI Act treats AI systems used for recruitment, performance evaluation, task allocation and monitoring of workers as high-risk. Recent reporting indicates that lawmakers agreed to delay the application of the high-risk obligations from 2 August 2026 to 2 December 2027, with formal adoption steps following the political agreement (source: employment law firm briefings from Fisher Phillips and DLA Piper). Confirm the current status before you publish, because timelines have moved repeatedly. Importantly, the Act's prohibited-practice rules have applied since February 2025 and were not part of the delay, so employers should check that their tools do not stray into banned territory such as inferring emotions in the workplace.
Data protection law. In the UK and EU, systematic monitoring of employees generally triggers a duty to assess the risk to individuals before deploying, typically through a data protection impact assessment, along with transparency, proportionality and purpose-limitation obligations. Rules on decisions made solely by automated means add further constraints on using scores to discipline or dismiss.
Other jurisdictions. Several US states and cities have introduced rules on automated employment tools, mainly in hiring. Requirements vary widely, so take local legal advice.
The practical takeaway: if a tool scores people and influences their pay, progression or job security, assume regulators will treat it as high-risk, and design your governance accordingly even where enforcement dates have slipped.
Ten Questions to Ask Before Buying Enterprise Performance Management Software
Whether you are evaluating enterprise performance management software, enterprise HRIS systems or a module inside enterprise HR software, put these questions to every vendor in writing.
What exactly is measured, and why is it a valid proxy for quality? Demand evidence, not marketing claims.
What are the false-positive and false-negative rates on people like ours? Ask for results on your accents, dialects, roles and languages.
What does a "confidence score" mean mathematically? If it is not calibrated, do not display it as a percentage.
Can employees see their own scores, evidence and comments before managers do? Transparency reduces fear.
Is there a documented human-review step before any consequence? Flags should trigger a conversation, never an automatic penalty.
How do employees challenge a score? An appeals route is essential.
What data is retained, for how long, and who can access it? Minimise retention.
Which models process our data, and where? Understand sub-processors and data flows.
How does the system handle disability, neurodiversity, stammering and non-native speech? Test for disparate impact.
Can we switch off features? Modular deployment lets you keep coaching tools and drop punitive scoring.
If a vendor cannot answer questions 2, 3 and 9 clearly, treat that as a finding.
How to Roll Out AI Performance Tools Without Breaking Trust
Buying the right product is half the job. Implementation decides whether people experience help or surveillance. The following phased plan is an original framework drawn from the failure modes above.
[IMAGE 5: horizontal 90-day rollout timeline with four phases] Alt text: Four-phase 90-day rollout plan for responsible AI performance tools: consult, pilot, review, scale Caption: A staged rollout builds trust and catches errors before they reach careers.
Days 1 to 30: Consult and define. Involve employees and their representatives before configuration. Write down the purpose in one sentence and remove any metric you cannot tie to it. Complete the impact assessment. Decide in advance what the system will never be used for, such as automatic discipline.
Days 31 to 60: Pilot with shadow scoring. Run the tool silently on a small group. Compare its flags with trained human judgments and publish the false-alarm rate internally. Adjust rubrics, remove weak proxies and test on diverse voices.
Days 61 to 75: Coaching-first launch. Show scores to employees first. Frame the tool as a coaching aid. Give people a way to annotate or dispute flags. Track sentiment, absence and attrition alongside quality metrics, because stress shows up in those numbers before it shows up in surveys.
Days 76 to 90: Review and decide. Look at outcomes, not activity. Did learner results or customer outcomes improve? Did complaints fall? If the answer is unclear, do not scale. Set a sunset date so the system must earn renewal.
Connect performance tools to the wider stack thoughtfully. Enterprise workflow automation software can route flags to coaches and log every human decision, which creates the audit trail regulators expect. Meanwhile enterprise recruiting software and hiring analytics face the same high-risk scrutiny as monitoring tools, so apply the same ten questions there too.
The Business Case Against Over-Monitoring: A Simple Cost Model
Executives approve monitoring tools because the savings are visible: fewer manager hours spent on observation, faster detection of problems. The costs are harder to see, so they rarely appear in the business case. Here is a simple model you can adapt. All numbers below are assumptions for illustration, not figures from Multiverse or any real company.
Imagine an organisation with 200 skilled instructors, trainers or analysts. Assume the average replacement cost of a departing skilled employee (recruiting, onboarding, lost productivity while a new hire ramps up) is around 50% of annual salary, and the average salary is 40,000 in your local currency.
Replacement cost per departure: 20,000
Baseline annual attrition: 12%, or 24 people, costing 480,000 a year
If intrusive scoring pushes attrition up by just 4 percentage points, that is 8 extra departures, costing an additional 160,000 a year
Now compare that with the savings. If the tool saves each of 10 managers around 3 hours a week of observation time at an assumed 30 an hour, the annual saving is roughly 46,800. In this example, a four-point rise in attrition wipes out the observation savings more than three times over, before counting licence fees, legal review, sick leave or damage to learner outcomes.
The model has three uses. First, it shows why leaders should track attrition, absence and sentiment as formal success metrics for any monitoring tool, not just quality scores. Second, it explains why the cheapest-looking option, full coverage by software, is not always the cheapest overall. Third, it gives you a break-even test before you buy: ask what increase in attrition would cancel the projected savings, then decide whether you are confident the rollout will stay below it.
There is also a quality effect that the spreadsheet cannot capture. When people believe every sentence is being scored, they tend to move toward safe, scripted delivery. In a teaching or coaching context, that can reduce the natural responsiveness, humour and improvisation that make sessions effective. A tool that improves average compliance while flattening the best performers may look successful on a dashboard and still harm the customer or learner experience.
If you want a rule of thumb, use this: any monitoring feature should be able to show a plausible path to better outcomes for the people being served, not just a cheaper way to supervise the people doing the serving.
What This Means for Workers and Managers
If you are a worker under AI scoring, ask what is measured, request your own data, keep notes of disputed flags, and raise concerns collectively. Where you are in a union or works council, bring it to them. Stress from surveillance is a legitimate workplace health issue, not personal weakness.
If you are a manager, treat every flag as a hypothesis, not a verdict. Read the underlying evidence before you speak to anyone. Remember the false-alarm arithmetic: most months, most good people will trip something. Your credibility depends on how you respond when the machine is wrong.
If you lead the business, remember that the teachers at the centre of this story were reportedly praised for their skill even as they were being scored. Retention of skilled instructors is a competitive asset. A tool that drives them out to save review costs is a false economy.
The Bigger Irony
The reporting highlights an uncomfortable contrast. The company's co-founder is a prominent advocate for rapid AI adoption, arguing that workers will need retraining and that employers want AI-fluent workforces. Yet the staff delivering that retraining are the ones now reporting strain from AI oversight. That does not mean the technology is wrong, or that the firm's intentions are bad. It means adoption narratives and lived experience can diverge sharply, and the people closest to the tool usually spot the problems first. Leaders who want AI adoption to succeed should treat employee feedback as an early-warning system rather than resistance.
A note on transparency: the same coverage names Anthropic and OpenAI as providers of some models used by the firm. Model providers generally supply general-purpose capabilities, while the scoring rubric, thresholds, dashboards and management practices are chosen by the deploying organisation. That is exactly why buyer governance matters so much.
Frequently Asked Questions
Is AI employee monitoring legal? It depends on jurisdiction, purpose and implementation. In the EU and UK, monitoring must generally be transparent, proportionate and assessed for risk in advance, and employment-related AI is treated as high-risk under the EU AI Act, with obligations currently slated for late 2027. Take legal advice for your region.
What should enterprise performance management software measure? Prefer outcomes tied to real goals, such as learner results or customer satisfaction, over surface behaviours like filler words. Use behavioural signals only as prompts for coaching conversations.
Can AI improve performance management at all? Yes, when it supports coaching, gives employees their own feedback, and keeps humans accountable for decisions. Self-review and sampled review models are lower risk than continuous scoring.
How do I choose enterprise HR software that avoids these problems? Use the ten questions above, insist on transparency and appeals, run a shadow pilot, and choose modular products so you can disable punitive features.
What is the difference between enterprise HRIS systems and performance monitoring tools? HRIS systems are the system of record for employee data such as contracts, payroll and leave. Monitoring and scoring tools analyse behaviour. Connecting them increases both power and risk, so govern the integration carefully.
Where Employers Go From Here
The teacher-scoring story will not be the last of its kind. Vendors will keep selling continuous coverage, and boards will keep asking why they should not buy it. The organisations that come out ahead will be the ones that measure what matters, show people their own data, pilot before scaling, and keep a human accountable for every consequential decision.
If you are reviewing enterprise performance management software, enterprise human resources software or workflow automation for your people processes, start with governance, not features. The cost of a thoughtful rollout is small compared with the cost of a workforce that feels watched.
Place one commercial call to action here, for example: "Book a free 30-minute review of your workforce analytics and HR software stack" linking to your contact page. Replace with your real offer.
Comments