AutoML Beat a Hard Problem: How Bodyguard Cut Hate-Speech False Positives in Half

Key Takeaways

• Bodyguard, a French cyber-bullying protection app founded by Charles Cohen in 2017, built its hateful-comment detection model using OVHcloud’s AutoML platform rather than a custom-built machine learning pipeline, and the choice worked: detection accuracy rose from 80 to 90 percent, false positives fell from 6 to 3 percent, and the company reports deleting more than 2 million hateful comments in 20 months.
• AutoML automates the parts of building a machine learning model a data scientist would otherwise do by hand, choosing an algorithm, tuning its parameters, engineering features from raw data, which is precisely why a founder without a dedicated ML research team, exchanging nearly 80 pieces of metadata against a hard specification that the model’s error rate could not exceed 10 percent, could get a working, improving model into production rather than needing to build one from first principles.
• The specific problem Bodyguard set out to solve, distinguishing genuine hate from irony, sarcasm and humour, is a well-documented hard problem in natural language processing, not just a marketing framing. Academic research on sarcasm and irony detection consistently finds that standard sentiment and toxicity classifiers struggle to hold their accuracy on ironic text without additional contextual signal, which is exactly the gap Bodyguard’s 80 pieces of metadata, including reaction time and the percentage of upper-case letters, was built to help close.
• The regulatory backdrop has firmed up considerably since this kind of tool first mattered. Under the UK’s Online Safety Act, Ofcom’s illegal content codes of practice became enforceable on 17 March 2025 and its children’s safety codes require mitigation measures including age assurance from 25 July 2025, though Ofcom’s own consultation has explicitly declined to set a concrete cap on acceptable false-positive rates for automated moderation, leaving that judgement to individual services, Bodyguard’s own hard 10 percent ceiling included.
• AutoML’s real advantage here was speed and accessibility, not necessarily a higher performance ceiling than a bespoke model could reach. That’s a genuine, documented tradeoff in the wider AutoML literature, not a criticism specific to Bodyguard: automated pipelines lower the expertise barrier and cut time to deployment, while a hand-built model with continual retraining against evolving adversarial behaviour, exactly what content moderation faces as bad actors adapt, generally has more room to be pushed further over time. Bodyguard’s case study describes real, useful results within roughly two years of operation, not a claim that AutoML was the ceiling of what’s achievable.

The underlying problem Bodyguard set out to solve isn’t really a machine learning problem first, it’s a linguistics problem: telling the difference between a genuinely hateful comment and one that only reads as hateful out of context, sarcasm, dark humour, in-group banter. That’s a documented, actively researched hard case for automated systems, not something Bodyguard’s own marketing invented to sound impressive. What makes this case study useful for a UK C-suite is less the specific numbers, though they’re real, and more the choice behind them: Bodyguard built its detection model on an AutoML platform rather than assembling a bespoke machine learning pipeline, and that choice is worth understanding on its own terms, including where it helps and where it has real limits.

What Bodyguard actually built

Bodyguard is a real-time cyber-bullying protection app, and, per its published case study with OVHcloud, its predictive model now flags hateful comments with 90 percent accuracy, having deleted more than 2 million such comments over a 20-month period. The core technical challenge wasn’t just spotting an obviously abusive word, it was working out who a comment was actually aimed at and whether its tone was genuinely hostile or only read that way out of context, which meant the model needed an added layer capable of catching irony, sarcasm and humour rather than flagging them as hate. Cohen’s team built that model by feeding it nearly 80 separate pieces of metadata, among them how quickly a comment was posted after the content it responded to and what proportion of it was typed in upper case, against a specification that fixed the model’s acceptable error rate at no more than 10 percent.

“The technology also needed to be capable of detecting irony, sarcasm, and even humour. The predictive model created with the OVHcloud AutoML machine learning platform really helped me achieve this,” said Charles Cohen, founder and CEO of Bodyguard. He added: “I have never been a victim of cyber-bullying myself, but I grew up with everyone around me using social networks. I could see the harm caused by online hate. Cyber-bullying restricts freedom of expression, and this is what I struggled with when I was a teenager.” Cohen’s team used the platform to move detection accuracy up ten points, from 80 to 90 percent, while cutting the false-positive rate in half, from 6 to 3 percent, and relied on OVHcloud’s separate Logs Data Platform, paired with Grafana, to keep watch on infrastructure performance and key metrics as the system ran. Bodyguard itself is a real, established French company, founded in 2017 by Cohen, who built the underlying technology alone that year before the consumer app followed; by 2020 the company had reported around 75,000 app users, and OVHcloud’s own case study cites more than 40,000 users and 97 percent satisfaction as of July 2019.

What AutoML actually is, and what it isn’t

AutoML is a specific, well-defined category of tooling rather than a vague marketing term, it automates the parts of building a machine learning model that a data scientist would otherwise do by hand: selecting which algorithm to use, tuning that algorithm’s parameters, and engineering useful features out of raw data. That’s precisely why it suited Bodyguard’s situation. A founder without a dedicated in-house machine learning research team could still get a working model into production against a hard, defined specification, an error rate that could not exceed 10 percent, using nearly 80 pieces of metadata as inputs, without needing to build the underlying model-selection and tuning process from first principles.

What AutoML trades away for that accessibility is worth being honest about, and it’s a documented tradeoff in the wider technical literature, not a weakness unique to Bodyguard’s choice. Because AutoML automates model selection and tuning as a largely automated, multi-objective process, balancing accuracy, interpretability and speed without a practitioner hand-tuning every decision, it generally offers less fine-grained control than a bespoke pipeline built and continually refined by a specialist team. In adversarial domains specifically, where the people generating hateful content actively adapt their language to evade detection, a hand-built model with ongoing, deliberate retraining against evolving behaviour typically has more room to be pushed further over time than an automated pipeline that isn’t being continually retrained against the same moving target. None of that makes Bodyguard’s approach wrong, the results reported, a ten-point accuracy gain and a fifty percent relative cut in false positives, are real and useful achievements within roughly two years. It does mean AutoML should be read here as the tool that got Bodyguard to a strong, deployable result quickly, not as evidence that it represents the highest ceiling achievable on this specific, genuinely difficult problem.

The problem is genuinely hard, not just difficult to market

It’s worth taking Bodyguard’s stated challenge seriously rather than treating it as a founder’s framing of an easy problem to sound impressive. Detecting sarcasm and irony in text is an active, specifically named research area in natural language processing, and the consistent finding is that standard sentiment and toxicity classifiers struggle to hold their accuracy when text is ironic, because the words used often imply something close to the opposite of their literal meaning, a signal that requires context most classifiers don’t have access to by default. A peer-reviewed study specifically on sarcasm and irony classification applied to cyberbullying detection reinforces that this isn’t a hypothetical edge case for a tool like Bodyguard’s, it’s close to the central technical challenge the product exists to solve. The 80 pieces of metadata Bodyguard’s model draws on, including how quickly a comment was posted after the original content and how much of it is in upper case, are the kind of contextual proxy signal this line of research suggests is needed to close that gap, though the specific feature set is Bodyguard’s own and not something the wider literature prescribes directly.

The regulatory ground has shifted underneath this kind of tool

Bodyguard’s technology predates the UK’s current regulatory framework for online harms, but that framework is now directly relevant to any company building detection tools like it. Under the Online Safety Act, Ofcom’s codes of practice on illegal content became enforceable on 17 March 2025, covering seventeen categories of illegal harm including hate offences, with risk assessments due the day before. Separately, Ofcom’s Protection of Children codes, published in April 2025, require mitigation measures, including “highly effective” age assurance, from 25 July 2025. It’s worth being precise about what the Act does and doesn’t require here: its provisions on proactive automated detection are narrowly scoped, largely to child sexual abuse material, terrorism content and intimate image abuse, rather than a general legal mandate for AI-driven detection of harassment or hate speech of the kind Bodyguard targets.

On the specific tension Bodyguard’s own 10 percent error-rate specification was designed to manage, over-removal versus under-removal, Ofcom’s own consultation process has explicitly declined to set a concrete cap on acceptable false-positive rates for automated moderation tools, stating that the impact will depend on the service in question, leaving each platform to set and justify its own threshold rather than imposing one centrally. Commentary from legal and digital rights analysts has flagged that the Act’s free-expression safeguard is a comparatively weak “have regard to” duty rather than a binding standard, and that over-moderation carries its own real risk to legitimate speech. Bodyguard’s own hard ceiling, an error rate that could not exceed 10 percent, reads as a company setting a deliberately tighter internal standard than regulation currently requires, worth noting as a genuinely responsible design choice rather than a coincidence of the numbers looking tidy.

Why this problem is worth solving at the scale Bodyguard describes

The scale of the underlying problem gives Bodyguard’s figures real weight rather than treating them as an isolated success story. Ofcom’s own Children and Parents research, published in 2024, found that 32 percent of UK children aged 8 to 17 have experienced online bullying, with the specific behaviours measured affecting girls somewhat more than boys in that survey. Set against that backdrop, a tool reporting more than 2 million hateful comments deleted in 20 months, with a measured, improving accuracy rate rather than a vague claim of success, is addressing a documented, widespread problem rather than a manufactured one.

What a UK C-suite should take from this

The lesson isn’t that AutoML is always the right choice, or always a compromise. It’s that AutoML solved a specific, real problem for Bodyguard, getting a founder without an in-house ML research function to a working, improving, tightly specified model quickly, and that this is a genuinely different value proposition than the one a bespoke, continually hand-tuned pipeline offers to an organisation with the resources to build and maintain one. For a board evaluating a similar build, the honest question isn’t which approach is better in the abstract, it’s whether the organisation has, or is willing to build, the specialist capacity a bespoke pipeline requires to outperform AutoML over time, and whether the problem itself, like Bodyguard’s, is well enough defined by a clear specification, an error rate that cannot exceed a stated number, that an automated pipeline can be pointed at it and trusted to deliver.