Innocent Words, Harmful Data: How Outsourced Review and Regional Blind Spots Let Coded Manipulation Slip Into AI Pipelines

This paper argues that AI training data pipelines have blind spots because the people reviewing the data (low-paid contract workers using machine translation) can't catch harm that's encoded in cultural context, regional dialect, or coded language rather than in obvious keywords. The proposed fix is a two-tier system: expert trios (a Historian, a Political Scientist, and a Local Linguist) overseeing contract annotators per language/region, with independent auditing by universities and government bodies.

The underlying concern is real. Anyone who has followed reporting on the data annotation industry knows that reviewers work under speed quotas, often lack domain expertise, and routinely use machine translation to evaluate languages they don't speak. The observation that machine translation flattens coded language into clean-sounding prose, causing reviewers to miss culturally embedded harm, is a legitimate point. The case studies (coded sex-trade listings using a "wife voice" register, Western-centric hate-speech taxonomies applied to non-Western contexts, conflict-zone dialect drift across borders) illustrate failure modes that are plausible and that people working in content moderation would recognize.

However, the paper has fundamental problems with evidence, rigor, and fit to the hackathon's theme.

There is no empirical work here. The paper states this directly: "We do not execute code or train neural networks." There are no experiments, no measurements, no data collection, no statistical analysis. The case studies are described as "real-world failure modes," but none includes a citation to a documented incident, a dataset, a specific model, or a specific annotation platform. The "wife voice" case, the Sudan dialect case, the pharaonic incest case: are these things that actually happened in a specific pipeline, or are they hypothetical scenarios the author constructed to illustrate the argument? The paper doesn't say, and the absence of any sourcing makes it impossible to evaluate whether these are documented failures or plausible but invented examples. For a research submission, even a hackathon one, the difference matters.

The connection to secret loyalties is thin. The paper references Lamerton & Roger and Hubinger et al. to establish that hidden behaviors persist through safety training, then argues that the data pipeline is the primary defense. But the case studies are about general data quality failures (missing PII, culturally inappropriate content passing review, machine translation errors), not about the deliberate or inadvertent installation of loyalty to a specific principal. A reviewer missing coded sex-trade language because Google Translate sanitized it is a content moderation failure, not a secret loyalty. The paper gestures at the connection in Case 4 (state-sponsored actors injecting balanced-sounding propaganda) but doesn't provide evidence that this has happened or that the proposed governance framework would catch it. The hackathon's framing is about models being trained to covertly serve a named principal, and most of what this paper discusses is about general training data contamination.

The proposed solution (expert trios plus institutional auditing) is reasonable as a governance recommendation but is presented without any analysis of feasibility, cost, or effectiveness. How many language/region combinations would need coverage? What would it cost? How would you recruit enough historians and political scientists with the specific expertise needed? How would you handle the hundreds of languages and thousands of dialects that major models train on? How would university auditors access proprietary training data given trade-secret concerns? The paper lists these as limitations ("institutional coordination," "resource allocation," "jurisdictional conflicts") but doesn't engage with them substantively. As presented, the framework is an aspiration rather than a workable proposal.

The writing is clear and the structure is logical, but the paper relies heavily on formatting (flowchart-style ASCII diagrams, bolded subheadings, nested bullet points) that adds visual weight without adding analytical depth. Several passages restate the same point in slightly different words across sections. The paper would be stronger if trimmed to half its length and supplemented with even one concrete, sourced example of a documented pipeline failure.

On dual-use and responsible disclosure, there are no concerns since the paper contains no code, no trained models, and no operational details that could enable harm.

Overall: It identifies a genuine and important problem (cultural and linguistic blind spots in data annotation) but presents it as a conceptual argument without empirical evidence, with a weak connection to the hackathon's specific theme of secret loyalties, and with a proposed solution that lacks feasibility analysis. The case studies would be compelling if sourced; as presented they read as hypotheticals.

The author raised some interesting issues regarding flattening of language and adapting models to local contexts that are definitely worth exploring. I rated the submission low primarily on the basis that I was not convinced there was a major safety concern. An LLM's ability to keep up with the most recent sudanese hate speech doesn't seem like a particularly pressing safety concern. I also don't see much application to the larger hackathon focus on secret loyalties.