OpenAI has a goblin problem.

Instructions designed to guide the behavior of the company’s latest model as it writes code have been revealed to include a line, repeated several times, that specifically forbids it from randomly mentioning an assortment of mythical and real creatures.

“Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user’s query,” read instructions in Codex CLI, a command-line tool for using AI to generate code.

It is unclear why OpenAI felt compelled to spell this out for Codex—or indeed why its models might want to discuss goblins or pigeons in the first place. The company did not immediately respond to a request for comment.

OpenAI’s newest model, GPT-5.5, was released with enhanced coding skills earlier this month. The company is in a fierce race with rivals, especially Anthropic, to deliver cutting-edge AI, and coding has emerged as a killer capability.

In response to a post on X that highlighted the lines, however, some users claimed that OpenAI’s models occasionally become obsessed with goblins and other creatures when used to power OpenClaw, a tool that lets AI take control of a computer and apps running on it in order to do useful things for users.

“I was wondering why my claw suddenly became a goblin with codex 5.5,” one user wrote on X.

“Been using it a lot lately and it actually can't stop speaking of bugs as ‘gremlins’ and ‘goblins’ it's hilarious,” posted another.

The discovery quickly became its own meme, inspiring AI-generated scenes of goblins in data centers, and plug-ins for Codex that put it in a playful “goblin mode.”

AI models like GPT-5.5 are trained to predict the word—or code—that should follow a given prompt. These models have become so good at doing this that they appear to exhibit genuine intelligence. But their probabilistic nature means that they can sometimes behave in surprising ways. A model might become more prone to misbehavior when used with an “agentic harness” like OpenClaw that puts lots of additional instructions into prompts, such as facts stored in long-term memory.

OpenAI acquired OpenClaw in February not long after the tool became a viral hit among AI enthusiasts. OpenClaw can use any AI model to automate useful tasks like answering emails or buying things on the web. Users can select any of various personae for their helper, which shapes its behavior and responses.

OpenAI staffers appeared to acknowledge the prohibition. In response to a post highlighting OpenClaw’s goblin tendencies, Nik Pash, who works on Codex, wrote, “This is indeed one of the reasons.”

Even Sam Altman, OpenAI’s CEO, joined in with the memes, posting a screenshot of a prompt for ChatGPT. It read: “Start training GPT-6, you can have the whole cluster. Extra goblins.”

A movement is underway to classify heavy social media use as a form of addiction. Advocacy groups, plaintiffs’ attorneys, and a growing number of lawmakers are treating the proposition as settled science. In a landmark California trial in early 2026, a jury found Meta and Google negligent for designing platforms that allegedly caused mental health harm, awarding $6 million in damages. 

In Congress, the Kids Online Safety Act, a bill that would require social media platforms to prevent specified harms to minors, including “compulsive usage,” and to disable addictive product features by default, has advanced out of committee in both chambers. Additional bills would ban children under age 16 from using social media entirely, require age verification at the app store level, and expand the Children’s Online Privacy Protection Act to cover minors up to age 17. Australia has enacted an outright ban on social media for children under 16 years old.

The pace of legal and legislative action suggests a society that has made up its mind. But the underlying science has not. The research base on social media addiction remains fragmented, methodologically inconsistent, and far from the kind of consensus that would ordinarily justify the regulatory and legal apparatus now being built around it. As we have argued in our earlier analysis of how the American health care system rewards psychiatric overdiagnosis, when diagnosis is subjective and payment depends on diagnosis, the system will predictably expand the boundaries of illness. Social media addiction is poised to become the next case study in that dynamic, with consequences that extend well beyond health care spending into the domains of free speech, privacy, and innovation.

Addiction, Dependence, and Habit Are Not the Same Thing

Before asking whether social media is addictive, it is worth clarifying what addiction actually means, because politicians, journalists, and even some clinicians routinely misuse the term. As one of us has previously argued, the conflation of addiction with dependence and habit distorts both public understanding and public policy.

These three concepts describe very different things. Dependence is a physiological adaptation in which abruptly stopping a substance produces withdrawal symptoms. It is common and, by itself, unremarkable. Anyone who has ever quit coffee cold turkey and spent the next two days with a splitting headache has experienced caffeine dependence. Patients who take certain antidepressants, benzodiazepines (e.g., Valium), antiepileptic drugs, or beta blockers for extended periods develop physical dependence as well. If they stop abruptly, they will experience withdrawal. In some cases, withdrawal can be fatal. Yet no one would claim that patients taking beta blockers long term for high blood pressure or antiepileptic medications for a seizure disorder are addicted to those drugs.

A habit is something different still. Habits are behavioral patterns, often automatic, that people repeat because they find them pleasurable, comforting, or simply routine. Checking social media first thing in the morning, scrolling through a feed while waiting in line, or reaching for the phone out of boredom: These are habits. They may be unwise. They may waste time. They may even be difficult to break. But difficulty is not pathology. People also find it hard to stop snacking, binge-watching television, or hitting the snooze button. We do not diagnose these behaviors as diseases.

Addiction is distinct from both. The American Society of Addiction Medicine’s definition states: “Addiction is a treatable, chronic medical disease involving complex interactions among brain circuits, genetics, the environment, and an individual’s life experiences. People with addiction use substances or engage in behaviors that become compulsive and often continue despite harmful consequences.” Addiction can involve substances or activities. For example, the American Psychiatric Association’s Diagnostic and Statistical Manual of Mental Disorders (DSM) classifies gambling disorder as an addiction. 

The defining feature is compulsive behavior: repeated engagement despite clear harm to relationships, finances, or health. This pattern reflects changes in the brain’s reward and decisionmaking circuits that erode self-control. But people with addiction do not completely surrender agency; they are not zombie-like automatons at the mercy of a substance or activity. Even in the grip of addiction, many will avoid use in certain settings, delay it when consequences are immediate, or respond to incentives—evidence that agency, though impaired, is still intact.

When the term “addiction” is applied loosely to heavy social media use, it skips over the crucial middle category of habit and confers a clinical gravity that the evidence does not support. Someone who checks Instagram too often or finds it hard to put down TikTok almost certainly has an unhealthy habit. Calling it an addiction equates that behavior with the compulsive, life-destroying patterns seen in substance use disorders.

This distinction matters because “addiction” is not a neutral word—it has specific consequences enshrined in policy. Once applied, it unlocks diagnosis codes, insurance payments, treatment industries, lawsuits, and regulation. We don’t create Medicaid billing categories for habits. We don’t pass federal laws to shield children from habits. Label something “addiction,” and the entire policy engine comes to life. That label should follow the science, not lead it.

What Does the Research Actually Show?

Much of the existing research on social media and mental health suffers from serious methodological limitations. The majority of studies rely on self-reported measures administered at a single point in time. They can identify correlations between heavy social media use and negative mental health outcomes, but they cannot establish the direction of causation. It is equally plausible that individuals already experiencing depression, anxiety, or social isolation turn to social media as a coping mechanism rather than social media causing those conditions.

A 2024 systematic review and meta-analysis published in JAMA Pediatrics analyzed 143 studies examining the effects of social media use on mental health among over one million adolescents worldwide. Overall associations were small and inconsistent across studies and often confounded by other factors such as personality and social support. The evidence base being cited to justify sweeping policy interventions is built on faulty scientific ground, as one of us recently argued in the Washington Post.

None of that means excessive social media use cannot be harmful for certain individuals. Some people undoubtedly experience significant distress and functional impairment related to their online habits. But the question of whether a behavior causes harm in some people is different from the question of whether it constitutes a discrete clinical disorder, and the latter question is far from resolved.

Why We Should Not Trust the Diagnostic Authorities to Get This Right

Proponents of recognizing social media addiction as a disorder often point to the DSM and the International Classification of Diseases (ICD) as the bodies that will eventually resolve the question. But formal inclusion in these manuals would not settle the science. It would settle the payment. And the track record of these manuals should inspire caution, not confidence.

Formal classification matters because of what it triggers financially. In the American health care system, a diagnosis unlocks reimbursement. A recognized social media addiction diagnosis would trigger insurance coverage under the Mental Health Parity and Addiction Equity Act, which requires health plans, including Medicaid managed care, to cover behavioral health services at parity with medical and surgical services. Under fee-for-service billing, providers increase revenue by increasing the volume of services delivered, and subjective diagnostic criteria provide the discretion to do so. The system fixes the price of a service but introduces no effective mechanism for governing whether the service was clinically necessary. 

The DSM has progressively broadened the boundaries of psychiatric illness over successive revisions, often without corresponding improvements in diagnostic precision. Its fifth edition collapsed previously distinct autism categories into a single spectrum elastic enough to encompass both nonverbal children requiring constant care and socially awkward adolescents who prefer solitude. It loosened ADHD criteria, allowing symptom onset as late as age 12 rather than requiring it by age 7, and reduced the symptom threshold for adults. Generalized anxiety disorder requires only that worry be “excessive” and cause “clinically significant distress or impairment,” judgments that depend entirely on a clinician’s interpretation of where normal worry ends and disorder begins. Each revision has expanded the population eligible for diagnosis and, with it, the population eligible for treatment and reimbursement.

As we documented in our recent analysis of Medicaid-funded autism therapy, the broadening of autism spectrum criteria, combined with Medicaid’s open-ended reimbursement structure, produced an explosion in spending on applied behavior analysis therapy that far outpaced any plausible change in the actual prevalence of disabling autism. The broadening of ADHD criteria produced a parallel surge in stimulant prescriptions. In each case, the combination of subjective diagnosis and financial incentives that reward diagnosis pushed the boundaries of illness outward.

Social media addiction, if formalized, will follow the same trajectory. A social media addiction rehabilitation industry is already emerging: Specialized retreats, counseling programs, and screen-time management apps are marketing themselves to anxious parents and burned-out professionals. That industry will expand dramatically the moment a formal diagnosis is established—a new special interest group, funded in significant part by taxpayer dollars and private insurance premiums. 

We have already seen this dynamic play out when unhealthy habits become medically pathologized. The inclusion of “gaming disorder” in the ICD-11 in 2019 is a cautionary example rather than a reassuring precedent. A large group of scholars published an open letter opposing the classification, warning that it rested on a low-quality evidence base, that the diagnostic criteria leaned too heavily on substance use and gambling frameworks without adequate validation for behavioral contexts, and that official classification would generate a “tsunami of false positive referrals to treatment.” The scholars were particularly concerned that premature classification would pathologize ordinary recreational activity and cause significant stigma. That gaming disorder made it into the ICD-11 despite these objections is not evidence that the process works. It is evidence that diagnostic classification is driven as much by political and institutional momentum as by scientific rigor. Anyone who believes that social media addiction will receive more careful treatment from these same institutions is not paying attention.

What Comes Next

This pattern is by now familiar. Subjective diagnostic criteria and financial incentives that reward diagnosis have repeatedly produced policy responses that were disproportionate, costly, and harmful. The social media addiction debate is following the same trajectory. If we do not insist on scientific rigor before enshrining a diagnosis in law and policy, we will once again find ourselves managing the consequences of a premature consensus.

The question is not whether social media can be used in unhealthy ways. Of course it can. The question is what happens when we reclassify those behaviors as a medical disorder before the science supports that classification. When diagnosis is subjective and incentives reward diagnosis, the boundaries of illness expand. More diagnoses generate more treatment, more spending, and more regulation—along with greater government intrusion into choices that were once considered matters of personal judgment.

That reclassification also invites litigation. Once courts accept the premise of “addiction,” lawsuits will pressure platforms to alter or restrict lawful content and design features, often through settlement agreements negotiated outside the legislative process. Lawmakers, responding to a perceived epidemic, will layer on additional restrictions—limiting access, mandating intrusive age verification, and expanding regulatory oversight of online speech. As our colleagues at Cato have argued, these interventions threaten free expression and innovation while doing little to address the underlying concerns about children’s welfare and risk undermining online speech and privacy for users of all ages. The costs will be measured not only in dollars but also in diminished speech, eroded privacy, and the further medicalization of ordinary human behavior.

When the science is unsettled and the incentives to expand diagnosis are strong, we must be cautious. At bottom, this is a question of restraint.

We should be especially careful before turning a widespread human behavior into a medical disorder, because once we do, the consequences will extend far beyond the people we are trying to help.

London's Metropolitan Police Service (MPS) has survived a legal challenge that attempted to curb its rollout of live facial recognition (LFR) technology across the capital.

The challenge was brought against the Met by civil liberties organization Big Brother Watch, which was representing Shaun Thompson, an anti-knife crime campaigner and youth worker who was falsely identified as a criminal suspect by LFR cameras in Croydon.

Big Brother Watch supported Thompson's case, which argued that the technology violated his rights to privacy under articles 8, 10, and 11 of the European Convention on Human Rights (ECHR).

The UK's High Court concluded this week that LFR technology itself does not violate any of the ECHR's aforementioned articles, and found that Thompson's personal rights to privacy were not infringed.

Presiding over the case, Lord Justice Holgate and Mrs Justice Farbey considered both the UK's incorporation of the ECHR into domestic law (the Human Rights Act 1998) and Strasbourg's, but found that the Met's LFR policy satisfied the requirement for being in accordance with and prescribed by the law.

In short, the justices found the Met's planned use of LFR is legal and does not violate the human rights of Britons who are subjected to it.

Sir Mark Rowley, commissioner of the MPS, described the judgment [PDF] as "a significant and important victory for public safety."

"The courts have confirmed our approach is lawful. The public supports its use. It works. And it helps us keep Londoners safe. The question is no longer whether we should use Live Facial Recognition - it's why we would choose not to.

"Technology is advancing at record speed, and policing cannot afford to stand still – criminals won't. Facial recognition is transformational for policing. Government and Parliament will want to carefully consider how they continue to enable, rather than over‑regulate, the use of technologies that help us prevent crime and protect the public as proven today."

Silkie Carlo, director at Big Brother Watch, labeled the High Court's judgment "disappointing." As for Thompson, he plans to appeal the decision.

"I've considered the court's judgment today and decided to appeal it to protect Londoners from facial recognition being used for mass surveillance and leading to situations like mine, where I was misidentified, detained, and threatened with arrest," Thompson said.

"No one should be treated like a criminal due to a computer error.

"I was compliant with the police, but my bank cards and passport weren't enough to convince the police the facial recognition tech was wrong. It's like stop and search on steroids. It's clear the more widely this is used, the more innocent people like me risk being criminalized."

A hot topic

Police use of LFR in the UK is a fiercely debated topic. Law enforcement officials insist it is an invaluable tool to protect public safety, while privacy proponents argue it represents a severe surveillance overstep.

The Met, meanwhile, claims the tech has led to 2,100+ arrests since 2024, saying a quarter of these (24 percent) were related to violent crimes against women and girls. It also claims more than 100 sex offenders were arrested off the back of LFR, and the identifications potentially prevented many more sex attacks against vulnerable children.

In their unwavering support for the technology, police forces often spout the results from independent safety tests to which LFR systems are subjected before they are deployed.

The National Physical Laboratory carries out these assessments, and as The Register previously reported, the Met likes to frame the results in positive ways.

However, despite the Met claiming the technology is consistently performant across demographic groups, the false positive rates for Black people, including Thompson, are considerably higher than for any other group, and have been throughout various tests since at least 2020.

In the police's most recent annual review, it claimed low false positive rates of 0.0003 percent across a total of 3,147,436 faces it scanned across all deployments. But if you look at it in terms of the number of alerts LFR cameras specifically made (2,077), it rises to 0.48 percent. And of the false positives, 80 percent of them were made on Black people.

"Overall, the system's performance remains in line with expectations, and any demographic imbalances observed are not statistically significant," the report stated. "This will remain under careful review."

The UK government is approving wider deployments of LFR-equipped vans and permanent deployments despite the flaws, which in some cases are so significant that they are still preventing police forces from rolling them out. ®

Wolfgang Munchau is the Director of Eurointelligence and an UnHerd columnist.

EuroBriefing