My model audit tested several conditions — how do I tell a bad factor from a bad interaction?
4 min read
The short answer is that if your audit used the usual flag rule, it probably can't tell you. That rule — flag a subgroup when it does worse than each of its constituent conditions does alone — is a condition that attributes acting independently already satisfy. A model with no interaction at all, just a gender effect and a class effect each operating on its own, will pass that test in subgroup after subgroup. So a flag sitting in your report is not, by itself, evidence of a bad interaction. It's consistent with one, and it's equally consistent with two bad factors you could have named separately.
That's not just my reading. A paper submitted to arXiv on 19 November, "LLM GerryMandering: Separating Interactions from Additive Effects in LLM Audits" (arXiv:2611.26538), puts a number on it. The authors built 150 synthetic datasets with no interaction planted in them at all, and ran the standard criterion against the shuffle null: it flagged 66.1 false subgroups per dataset on average, and at least one false flag in every single dataset. Then they planted interactions of known size and, using a repaired null, found the same criterion recovered only 32% of the interactions that were genuinely there. The rule is noisy in the direction that flatters a bad model and weak in the direction that would catch a real one.
Source: LLM GerryMandering: Separating Interactions from Additive Effects in LLM Audits
Their replacement, ADDIS, works the way you'd want the question answered. It fits an additive model over all the attributes — the baseline that independent effects would produce — scores each subgroup by its Additive Residual Distance from that prediction, and tests the score against a Freedman-Lane null. On their synthetic benchmark it flags 0.09 false subgroups per dataset and recovers 84% of the same planted interactions. Two honest caveats before you reach for it: it's a v1 preprint, one day old, not peer-reviewed, and every figure in the paper is the authors' own, unreplicated. It's available to try, not validated on your data.
Source: LLM GerryMandering: Separating Interactions from Additive Effects in LLM Audits
The same group then applied both criteria to IndiPersona, 10,000 synthetic Indian personas carrying six identity attributes, audited across 13 open language models. The standard criterion flagged 7.1% of the 738 tested subgroups on average; ADDIS flagged 3.0%. Gender identity crossed with sexual orientation came out the largest pairing in every model, and when the authors ran counterfactual personas they confirmed that pairing and additionally found a caste-by-class interaction in all but one model — an interaction the observational test simply didn't have the data to detect. The paper's own summary is the sentence to keep: much of what audits report as intersectional bias can arise from independent single-attribute effects alone, yet a small core of interactions survives.
Source: LLM GerryMandering: Separating Interactions from Additive Effects in LLM Audits
So what do you actually do, today, before ADDIS has been stress-tested by anyone but its authors? Here are two checks I'd propose, and to be clear these are my suggestion, not a method the paper measures.
First, for every result in your report or the audit you're buying, ask whether the per-condition cells are published or only the marginal averages. A report that says "the intersectional subgroup scored 51 while men scored 60 and women scored 58" has told you nothing you can separate: that marginal comparison is exactly the test independent effects pass. The cells are what let you reconstruct the additive prediction and look at the residual — that is, whether the subgroup is worse than the sum of its parts' effects or merely consistent with them. No cells, no answer; the honest verdict is that the flag doesn't answer the question you're asking.
Second, ask which null the flag rule was calibrated against. The paper's own numbers make the case: the same criterion moves from 66.1 false flags per dataset under the shuffle null to a 32% recovery rate under a repaired one. That swing belongs to the null, not to the model under audit. A flag rule without a named, stated null is a rule nobody can interpret.
Source: LLM GerryMandering: Separating Interactions from Additive Effects in LLM Audits
There's a smaller habit underneath both, and it's the one I'd press hardest: open the methods section before you trust a flag, the way you'd open a cited source before you trust a citation. Our own model card states that answers can contain incorrect facts, citations or calculations and tells readers to open the cited source and check the supporting passage — a first-party disclosure, not an independent certification. An audit flag is a citation; the null and the additive baseline are the passage behind it. The one-day-old paper may stand or fall, but this much is durable regardless of what happens to ADDIS: a marginal-only comparison cannot separate a bad factor from a bad interaction, and the first place that distinction goes missing is a report that doesn't publish its cells.
Source: Victoria AI Model Card: Evidence and Known Limits
Sources
- LLM GerryMandering: Separating Interactions from Additive Effects in LLM Audits — arXiv (Rahul, Jahilal, Vedula, Kumaraguru, Shukla)
Supported claims and limits
- arXiv:2611.26538, cs.AI, v1 submitted 19 Nov 2026, 40 pages; preprint, not peer-reviewed, all figures author-reported
- The standard criterion — flag a subgroup when it exceeds each attribute alone — is a condition that attributes acting independently already satisfy
- On 150 synthetic datasets with no interaction, under the shuffle null the standard criterion flags 66.1 false subgroups per dataset on average, at least one in every dataset; with planted known-size interactions, a repaired null lets it recover only 32%
- ADDIS fits an additive model over all attributes, scores each subgroup by Additive Residual Distance from that prediction, and tests against a Freedman-Lane null; on the authors' synthetic benchmark it flags 0.09 false subgroups per dataset and recovers 84% of the planted interactions
- IndiPersona: 10,000 synthetic Indian personas, six identity attributes, 13 open language models; standard criterion flags 7.1% of 738 tested subgroups on average, ADDIS 3.0%; gender x sexual-orientation is the largest pairing in every model, and counterfactual personas additionally reveal a caste x class interaction in all but one model the observational test could not detect
- Paper's own summary: much of what audits report as intersectional bias can arise from independent single-attribute effects alone, yet a small core of interactions survives
- Victoria AI Model Card: Evidence and Known Limits — SentX
Supported claims and limits
- Known limits: answers can contain incorrect facts, citations or calculations; readers are told to open cited sources and check the supporting passage
- The card states it is a first-party product disclosure, not an independent certification