Ethics, Quality Standards, and Expert Witness Practice

The International Association of Forensic Linguists' guidelines for practitioners set out the professional expectations for linguists acting as expert witnesses. The core expectation is that the expert's duty is to the court, not the instructing party. This is not unique to forensic linguistics. It appears in the guidelines of forensic pathologists, forensic accountants, and handwriting examiners. What makes it worth examining carefully in forensic linguistics is that the field is younger, less standardised, and more dependent on individual academic judgment than more established forensic disciplines.

The practical content of the duty to the court includes several specific obligations. The expert must disclose the limitations of their methods and the strength of their conclusions honestly. If the analysis found features pointing in the opposite direction from the instructing party's case, those must be disclosed, in the report and on the stand. The expert must not shade an opinion to strengthen a side's position. And the expert must be willing to change an opinion if cross-examination or additional evidence warrants it.

Daubert v. Merrell Dow Pharmaceuticals, decided by the US Supreme Court in 1993, restructured the rules for expert evidence in federal courts. The Court held that Rule 702 of the Federal Rules of Evidence requires trial judges to act as gate-keepers, evaluating the reliability of expert methods before they are put to juries. The Court identified four non-exclusive factors for that evaluation:

1. Whether the method has been empirically tested.
2. Whether it has been subjected to peer review and publication.
3. Whether there is a known or potential error rate and whether there are standards controlling the method's operation.
4. Whether the method is generally accepted within the relevant scientific community.

Kumho Tire Co. v. Carmichael (1999) extended Daubert's gate-keeping to all expert testimony, including technical and experience-based opinion, not only scientific evidence. For forensic linguists in US federal proceedings this means that authorship attribution methods, speaker profiling claims, or discourse analysis frameworks can all be challenged at a Daubert hearing before reaching the jury.

The older Frye standard (from Frye v. United States, 1923, in a case involving a systolic blood pressure deception test) requires only that the method be 'generally accepted' in the relevant scientific community. Frye is still applied in some US states. It is generally less demanding than Daubert for established methods but provides less protection against novel techniques that have achieved peer acceptance without rigorous empirical testing.

The leading case in England and Wales is R v Reed and Reed [2009] EWCA Crim 2698, a case involving DNA evidence. The Court of Appeal confirmed that expert opinion evidence is admissible if: it falls within a recognised body of expertise founded on research or experience; the witness has appropriate qualifications or experience; and the evidence is likely to be helpful to the jury. The English approach does not replicate Daubert's structured factor test, and trial judges have traditionally been more permissive about admitting expert evidence at the gate-keeping stage, relying instead on cross-examination and competing expert evidence to expose weaknesses.

The Law Commission's 2011 report on expert evidence in criminal proceedings recommended a more reliability-focused admissibility test closer to the Daubert model. Some of that thinking was incorporated into the Criminal Procedure Rules and Practice Direction amendments that followed, but English law still does not require the four-factor Daubert analysis for every expert. In practice, a forensic linguist's methods are more likely to be challenged through cross-examination or a competing expert's report than through a pre-trial admissibility ruling.

In civil-law jurisdictions (France, Germany, the Netherlands, India's civil proceedings), the court typically appoints the expert itself rather than receiving adversarial experts from each party. The court-appointed expert model changes the dynamic substantially. The expert's duty to the court is structurally easier to observe when the court is doing the appointing, but the model also reduces scrutiny because there is often no competing expert pressing for disclosure of limitations.

Daubert's third factor, the known error rate, is the most uncomfortable one for forensic linguistics. Most forensic linguistic methods do not have published, replicated error rate studies. This is partly a matter of the field's age and partly a matter of method: linguistics grew out of humanities scholarship, where the concept of a false-positive rate was not a natural analytical category. But it is also a structural problem. Many operational forensic tasks, such as attributing a short anonymous threatening letter to one of three suspects, do not resemble the clean test conditions that would permit rigorous error-rate measurement.

The best-studied area is computational authorship attribution. Studies using literary texts have achieved high accuracy in controlled conditions, but controlled conditions rarely match casework. Short texts, texts with deliberate style manipulation, texts written under stress, and texts in non-standard registers all reduce the performance of even the best computational methods. The honest position, and the one the IAFL guidelines push toward, is that a forensic linguist should be explicit about what validation exists for the method being used and what conditions that validation was established under.

Confirmation bias is the tendency to seek, interpret, and recall information in ways that confirm a pre-existing hypothesis. For a forensic linguist, the structural risk is real. An expert retained by the defence receives the case file with the defence theory visible. An expert retained by the prosecution receives it with the prosecution's framing. The analysis begins in a context that has already flagged one hypothesis as the working model. Ambiguous linguistic features, features that could support either side's case, are susceptible to being assimilated to the hypothesis the analyst arrived with.

Research on other forensic disciplines documents confirmation bias effects at operational level. Itiel Dror's studies of fingerprint examiners showed that the same examiner, presented with additional contextual information suggesting a match, changed a previous non-match decision at statistically significant rates. There is no reason to believe forensic linguists are structurally different. The methods are less standardised, the stimuli more complex, and the scope for motivated reasoning arguably larger.

• Blind analysis: the analyst should, where possible, not know which sample is the suspect's until the initial analysis is complete. This is structural, not optional.
• Competing-hypotheses protocol: explicitly test the hypothesis that the suspect did NOT write the text, not only the hypothesis that they did. Report both analyses.
• Mandatory disclosure of negative findings: features pointing against the instructing party's case must appear in the report, not only in the internal working notes.
• Peer review of the report: an independent colleague reviews the analysis before it is submitted to the instructing party. This is recommended in IAFL guidelines and required in some jurisdictions.

Science advances through peer review and replication. Forensic practice advances more slowly, in a context where case data is often confidential. This creates a tension: the best way to validate a method is to publish the casework data and let other experts test the analysis, but confidentiality restrictions, ongoing proceedings, and client privilege frequently prevent this. The result is that many forensic linguistic methods improve primarily through the accumulated private experience of practitioners rather than the public scientific record.

Some progress is possible within these constraints. Anonymised case studies, drawn from concluded proceedings with identifying information removed, allow methodological discussion without compromising parties. Academic publications that report on blind test sets, drawn from publicly available data, can establish error rates for specific tasks even if they do not cover operational conditions exactly. Language and Law / Linguagem e Direito, a peer-reviewed journal sponsored by the IAFL, publishes casework analyses, creating at least a partial public record.

The minimum transparency standard for an expert report, regardless of jurisdiction, is that the report should contain enough methodological detail that a qualified opposing expert can assess whether the analysis was sound. A report that says 'based on my experience, the texts were written by the same person' gives no opponent and no court enough to evaluate. A report that specifies which features were analysed, which comparison set was used, and which results cut against the conclusion gives the opposing expert something to engage with and the court something to weigh.