Nobody uses your platform the way you designed it I have spent the best part of 15 years studying how people commit crime, and I am the mother of children who live a good portion of their lives on platforms that were not built with them in mind. Both of those things make me wary of a sentence I hear constantly from trust and safety teams: "we already have controls in place for that."
Controls are designed for the user you imagined. The user you did not imagine is reading them too. For an offender, a safety control is not an obstacle so much as a piece of information. It tells them what the platform is looking for, where the platform is not looking, and what they need to do to stay below the line. Every keyword filter, age gate, report button and friend-suggestion setting is, in effect, a published description of your detection logic, and the people you built it to stop are its most attentive readers.
That is not cynicism. It is what the criminological evidence has shown for four decades, and it is now being confirmed at scale on the platforms children use every day.
Offenders adapt because adapting is cheap Situational crime prevention rests on a simple proposition: crime follows opportunity, so reduce the opportunity and you reduce the crime. Clarke's (1997) opportunity-reducing techniques have held up remarkably well. Guerette and Bowers' (2009) review of 102 evaluations of situational interventions found that displacement, the fear that crime simply moves elsewhere, was "the exception rather than the rule", and that where it did occur it was generally smaller than the gains made. That is the good news - controls do work.
But the same literature is equally clear that offenders are not passive. Ekblom (2017) describes prevention and offending as a co-evolutionary arms race in which each preventive move invites a countermove, from the thwarted burglar who returns with heavier bolt-cutters to the car thieves who found ways around keyless immobilisers within a few years of their introduction. Displacement is incomplete, but adaptation is constant. The offender who is blocked does not necessarily give up. More often, they learn.
The mistake platforms make is to read the first finding without the second. A control that reduces harm in its first year is filed as a solved problem, when what has actually happened is that the population it was aimed at has begun to study it.
What adaptation looks like on a platform Consider off-platforming. WeProtect Global Alliance (2023) Global Threat Assessment describes it as a common grooming technique: moving a conversation from the public space where a child was found into a private messaging app or an end-to-end encrypted environment "due to the lower risk of detection". The groomer uses the first platform's discoverability and the second platform's privacy, each working exactly as designed, to assemble a pathway that neither platform intended. The NSPCC (2025) analysis of 7,263 sexual communication with a child offences recorded by UK police forces in a single year found that Snapchat accounted for 40 per cent of the 2,111 cases where a platform was identified, and described "cycles of behaviours" in which perpetrators create multiple profiles and move children across services. Disappearing messages were built for privacy between friends. They are being used to destroy evidence.
Consider language. Steen, Yurechko and Klug (2023), interviewing TikTok creators, documented "algospeak": "seggs", "unalive", "le$bean", and a steady migration from altered spellings to gestures and visual cues as the moderation model caught up with each previous version. Their participants were largely benign users frustrated by clumsy, context-blind filtering, which is precisely the point. If sex educators can learn to route around a keyword classifier in weeks, there is no reason to think people with far stronger incentives are slower learners.
Consider platform choice. A survey of 20,592 self-reported users of child sexual abuse material, conducted by Protect Children and published by Ofcom (2026a), found that "many perpetrators avoid platforms with age checks or strict sign-up requirements, and actively choose platforms where they can maintain anonymity". Age assurance is a real control. It also works as a signpost, steering motivated offenders towards whichever service has not yet implemented it. And Ofcom's (2026b) demands that Snap Inc. , Meta and Roblox stop adult strangers contacting children, and stop prompting children to expand their networks, amount to an acknowledgement that ordinary growth features had become grooming infrastructure.
The message is the wrong unit of analysis Each of these controls fails in the same way: it inspects a moment, when the harm is a sequence. A keyword filter judges a message. An age gate judges a sign-up. A report judges a post. But grooming, coercive control, mule recruitment and sextortion are processes that unfold across dozens of exchanges, and the offender's entire craft consists of making each individual step look ordinary. Rapport, flattery, isolation, a suggestion to move somewhere quieter: no single message trips a wire, which is exactly why the adaptations above succeed. Change the word, change the platform, change the account, and a per-message system starts from zero every time.
Behavioural sequence analysis moves the unit of analysis from the message to the relationship. It asks not "is this word banned?" but "does this pattern of contact, over time, resemble the way exploitation is known to develop?" That is a much harder question to game, because to avoid it the offender would have to stop behaving like an offender. It is the approach we have built Tuteliq around, and it is why we train on sequences labelled by criminologists, psychologists and linguists rather than on lists of words.
Design for the user you did not imagine None of this is an argument against controls. The displacement evidence says they work; the adaptation evidence says they decay. The practical implication is that platform safety cannot be a feature you ship, only a capability you maintain against an opponent who reads your documentation more carefully than your users do. That means measuring whether harm has fallen or merely changed shape, assuming every new safety feature will be probed within weeks of launch, and anchoring detection in behaviour that offenders cannot easily abandon rather than vocabulary they can swap overnight.
If you run a platform that children use and want to see what longitudinal behavioural detection surfaces that your current filters miss, we would be glad to show you.
References Clarke, R.V. (ed.) (1997) Situational Crime Prevention: Successful Case Studies, 2nd edn. Guilderland, NY: Harrow and Heston.
Ekblom, P. (2017) 'Crime, situational prevention and technology: the nature of opportunity and how it evolves', in McGuire, M. and Holt, T. (eds) The Routledge Handbook of Technology, Crime and Justice. Abingdon: Routledge, pp. 353-374.
Guerette, R.T. and Bowers, K.J. (2009) 'Assessing the extent of crime displacement and diffusion of benefits: a review of situational crime prevention evaluations', Criminology, 47(4), pp. 1331-1368.
NSPCC (2025) Data shows how criminals are using private messaging platforms to manipulate and groom children, 13 November.
Ofcom (2026a) New research to support global efforts to tackle child sexual abuse and exploitation online, 18 March (reporting Protect Children's survey of 20,592 respondents).
Ofcom (2026b) Tech firms commit to stronger anti-grooming measures in response to Ofcom demands, 21 May.
Steen, E., Yurechko, K. and Klug, D. (2023) 'You can (not) say what you want: using algospeak to contest and evade algorithmic content moderation on TikTok', Social Media + Society, 9(3).
WeProtect Global Alliance (2023) Global Threat Assessment 2023.
Note on evidence: the Protect Children figures are self-reported by anonymous survey respondents and should be read as indicative of perpetrator perceptions rather than verified behaviour. The Steen et al. sample is 19 interviews with creators, so the algospeak findings are qualitative.