What does it mean that your AI scoring criteria are published?

It means the rules our system uses to judge speaking and writing are written down and available to read, rather than hidden. You can see what each CEFR level requires, how the descriptors map to a score, and where a human reviewer checks the outcome — before you buy, not after a dispute.

How do you prevent AI bias in the scoring?

We anchor every judgement to defined, published CEFR descriptors rather than to opaque preferences, key-match the objective skills so they are not open to interpretation, and add an editorial-review layer so a human can catch drift or an unfair edge case. Transparency is what makes bias auditable in the first place.

Can we contest or explain an individual result to a candidate?

Yes. Because the criteria are public and every result maps to a CEFR level, you can show a candidate exactly which descriptors they met and which they did not, and you can escalate a genuine edge case to our editorial review. A black-box score gives you nothing to explain or contest with.

Why We Publish Our AI Scoring Criteria (and Most Competitors Don't)

International English Test Editorial Team·6 Jul 2026·8

#AI scoring criteria#transparency#bias#black box#CEFR

The objection every buyer raises about AI-scored English tests

If you are evaluating an English assessment for hiring, admissions, or workforce development, you have almost certainly hit the same hesitation: can you trust a machine to score speaking and writing? It is the right question to ask. Reading and listening are easy to mark objectively, but productive skills are where judgement lives, and judgement is exactly where an opaque system can quietly go wrong.

Most of the market answers this question by asking you to take it on faith. The scoring engine is described as "advanced" or "proprietary," the results arrive with a confident number attached, and the actual rules are never shown to you. We take the opposite position. We publish our AI scoring criteria, and we pair them with a human editorial-review layer, because we think transparency is the only thing that turns a plausible score into a defensible one. This article explains the three fears buyers bring to AI scoring, how published criteria plus human review address each, and what that transparency actually lets you do once a result is in your hands.

The three fears behind "black box" scoring

Skepticism about AI-scored assessment usually breaks down into three distinct worries. They are worth separating, because a good answer looks different for each.

Bias. Does the model quietly favour certain accents, first languages, phrasings, or writing styles — and penalise others for reasons that have nothing to do with English ability?
Drift and inconsistency. Would the same answer score the same way today, next month, and after the model is updated? Or does the goalpost move without anyone telling you?
Opacity. When a result lands, can anyone actually explain why it is a B2 and not a C1? Or is the number simply asserted, with no reasoning you can inspect?

A single reassuring sentence on a sales page does not answer any of these. What answers them is being able to read the criteria yourself.

How transparency addresses bias

Bias thrives in the dark. If you cannot see what a system is rewarding, you cannot tell whether it is rewarding English proficiency or something correlated with it. Publishing the criteria does not by itself eliminate bias, but it makes bias auditable — and auditable bias is bias you can challenge, measure, and correct.

Our criteria are anchored to defined descriptors rather than to a hidden sense of what "good" sounds like. Objective skills are key-matched, so there is no room for interpretation to creep in on the parts of the test that have a right answer. And the editorial-review layer exists precisely so a human can catch a pattern the automated pass would miss. When the rules are on the table, a procurement team, an academic board, or a works council can inspect them and satisfy themselves that the system is judging the right thing.

How transparency addresses drift and consistency

Consistency is a promise, and promises need something to be checked against. Because our published criteria are tied to the CEFR scale — the same A1–C2 framework used across European education and employment — a score is not a free-floating number that can be re-tuned at will. It is a claim that a candidate has met a specific, stable set of level descriptors.

That anchoring is what keeps a B2 meaning the same thing across candidates, cohorts, and time. As an ALTE Associate Member, we hold ourselves to that external reference rather than to an internal one that only we can see. If you want the mechanics of how a raw performance becomes a level, we walk through it in inside our CEFR scoring model.

How transparency addresses opacity

Opacity is the fear that no one can explain the result — and it is the one that bites hardest in practice, because it surfaces at the worst possible moment: when a candidate, a hiring manager, or an auditor asks "why this score?" A black-box system leaves you with nothing but the number. Published criteria give you the reasoning behind it.

Every result maps back to descriptors you can read, and every judgement passes through a human editorial-review layer before it is treated as final. That combination is what makes a result something you can stand behind rather than something you simply hope is right. We set out the full approach in our scoring methodology, explained.

What published criteria actually let a buyer do

Transparency is not an abstract virtue. It changes what you can practically do with a result. Here is the contrast, stated plainly.

Buyer need	With published, human-reviewed criteria	With black-box scoring
Audit the method	Read the criteria and verify what is being measured before you commit	Take the vendor's word for it
Explain a result	Show the candidate exactly which descriptors they met	Assert a number with no reasoning
Contest an edge case	Escalate to editorial review against a known standard	No stable reference to appeal to
Defend a decision	Point to CEFR-anchored, documented criteria	Hope the result holds up under scrutiny

Each of those is a real moment in a real assessment programme. A candidate who narrowly misses a threshold will ask why. An internal auditor will want to know how the tool works before signing off. A regulator or an academic committee will expect the decision to rest on something documented. Published criteria mean you have an answer in every case, rather than a shrug.

This is also why transparency belongs before the pricing page, not after it. The trust question is the one that stalls a purchase. Answering it up front — by simply letting you read the rules — removes the largest single objection to adopting AI-scored assessment at all. For a fuller treatment of the standards a result has to meet, see what makes a result defensible.

Why most providers keep the criteria closed

If transparency is so obviously better for buyers, why is opacity the norm? Usually for reasons that serve the provider rather than the customer. Undocumented criteria are cheaper to maintain, harder to critique, and easier to change quietly. "Proprietary" can be a genuine engineering claim, but it is also a convenient way to avoid ever having to defend the actual rules.

The trouble is that a score you cannot inspect is a score you cannot fully rely on. The moment it is challenged — and in hiring or admissions, it eventually will be — an opaque system offers no way to explain itself. We would rather absorb the cost of publishing our criteria and inviting scrutiny, because scrutiny is what earns trust with buyers who have every reason to be cautious.

There is also a quieter cost to opacity that buyers feel later. When the rules are hidden, every downstream conversation inherits the same blind spot. A hiring manager cannot brief a candidate on what to improve. An admissions officer cannot reconcile a borderline result against their own rubric. A compliance reviewer cannot document how the decision was reached. Each of those gaps becomes your problem, not the vendor's, precisely at the point where you are trying to defend a decision to someone who was not in the room. Published criteria push that clarity upstream, so the people who rely on your assessment programme are working from the same visible standard you are — rather than reverse-engineering a number after the fact.

Transparency is the feature, not the disclaimer

AI scoring is not something to apologise for. Done well, it is faster, more consistent, and more scalable than manual marking. But "done well" has to mean something you can verify, not something you are asked to believe. Our answer is to publish the AI scoring criteria, key-match the objective skills, anchor every judgement to CEFR, and keep a human editorial-review layer in the loop — so that when someone asks how a result was reached, you can show them.

If you are weighing an assessment for your organisation and want to see exactly how the scoring works before you commit, explore our English assessment tests for companies.