There was a time when spotting a fake account on Twitter required little effort. A default egg avatar, zero followers, a username made of random letters and numbers, five posts all sent within the same hour — the signals were obvious enough that most users learned to recognise them without thinking. That era is over. The fake accounts operating on social platforms today are built differently, behave differently, and are designed specifically to pass the kinds of checks that used to be reliable.
Understanding why detection has become harder is the first step toward developing better instincts for it.
What Changed: The AI Factor
The most significant shift in the fake account landscape over the past two years is the widespread availability of generative AI tools for creating convincing profile content.
Profile pictures were historically one of the easiest tells. Stock photos were traceable through reverse image search; real photos borrowed from other accounts could be found at their source. AI-generated faces eliminated that vulnerability. A 2024 study published in the Journal of Online Trust and Safety identified over 1,400 Twitter accounts using faces generated by Generative Adversarial Networks (GANs) for their profile pictures, estimating a lower bound of 8,500 to 17,800 daily active accounts on the platform using this technique. Critically, the researchers also noted that newer diffusion-model images — the kind generated by tools like Midjourney and DALL-E — are even more convincing and render GAN-detection methods ineffective.
The same AI tools that generate faces also generate text. Where older bot accounts repeated the same phrases or posted clearly templated content, AI-generated posts can be varied, topically relevant, and stylistically consistent with a plausible human personality. A fake account can now maintain a coherent posting history across weeks or months, engage with trending topics in ways that feel authentic, and even produce replies that fit the specific context of a conversation.
The Behavioural Mimicry Problem
Beyond profile construction, the behavioural patterns that once distinguished fake accounts from real ones have become much harder to observe.
Early bot accounts were easily identified by activity patterns: posting hundreds of times per day, always at regular intervals, never interacting with replies, following thousands of accounts in bulk. Platform detection systems were trained on these patterns, and they became reliable signals for automated flagging.
Modern fake accounts are built to avoid these patterns deliberately. They post at irregular intervals, maintain realistic follower-to-following ratios, engage selectively with replies, and limit daily activity to volumes that fall within normal human ranges. A 2024 study testing eight major social platforms found that all of them failed to detect advanced AI-generated bot accounts during controlled experiments. Commercial anti-bot services fared only slightly better, with evasion rates measured at 44% and 52% against two widely-used tools.
This is not a solvable problem with existing detection infrastructure. The tools available to account creators have improved faster than the tools available to platform moderators.
Why Fake Accounts Look Real Now
Several specific techniques have made the visual and behavioural presentation of fake accounts convincingly human.
Aged account shells. Rather than creating accounts at the moment they are needed, operators purchase or build accounts years in advance and let them accumulate passive history — a few genuine-looking posts per month, occasional likes, gradual follower growth. By the time the account is activated for a campaign, it has a years-long history that passes most credibility checks.
Authentic-looking engagement networks. Fake accounts rarely operate in isolation. They exist within networks where accounts like, repost, and reply to each other, creating the appearance of genuine community engagement. A post that has been liked by 200 accounts feels more credible than one that has been liked by none — even if those 200 accounts are themselves fake.
Profile detail depth. Where early fake accounts had minimal bios and no linked content, modern ones include location information, website links, pinned posts, profile photos, banner images, and posting histories that span multiple topics. The effort invested in making a profile appear complete has increased substantially.
Realistic writing patterns. AI language models can generate posts that match the vocabulary, syntax, and even the typos that characterise human writing. Subtle grammatical inconsistencies — once a reliable signal for non-native or automated writing — are now deliberately introduced to make content appear more human.
What Still Works for Detection
Despite the improvements in fake account construction, several signals remain useful — though none are individually reliable enough to serve as a definitive indicator.
| Signal | What to check | Caveat |
| Profile photo | Reverse image search; look for symmetrical eyes and unusual texture around hairlines | Diffusion models now largely defeat visual detection |
| Account age vs. activity | Was the account dormant for years then suddenly active? | Aged shells are specifically designed to defeat this check |
| Follower network | Do followers have their own genuine-looking histories? | Coordinated networks mimic organic growth |
| Content consistency | Does the account post on a single topic with unusual frequency? | AI variation increasingly defeats this check |
| Link patterns | Do posts consistently link to the same domain or type of site? | More reliable than most behavioural signals |
| Engagement ratio | Does a low-follower account receive disproportionate engagement? | Suggests coordinated amplification |
The most reliable approach combines several signals rather than relying on any one of them. An account that raises questions on three or four of the above dimensions is far more likely to be inauthentic than one that triggers only a single concern.
The Verification Gap
Platforms have not kept pace with the evolution of fake account construction. Meta removed 1.4 billion fake accounts in Q4 2024 alone — a figure that reflects both the scale of the problem and the limits of automated detection, since accounts continue to be created faster than they can be removed.
For individual users, the practical implication is that platform verification signals — blue checkmarks, follower counts, engagement numbers — are less meaningful indicators of authenticity than they once were. An account can be verified, widely followed, and actively engaged with, and still be operating with inauthentic intent.
The burden of evaluation has shifted toward users. Knowing what signals to look for, and understanding why the obvious ones no longer work as reliably as they did, is now a necessary part of navigating social platforms in good faith.
