Text-Perfect AI Images Break Forensic Heuristics

TechCrunch tested OpenAI’s ChatGPT Images 2.0 and found it produces far more legible, believable text than earlier image generators. A single prompt produced a Mexican restaurant menu "that could immediately be used in a restaurant" without the gibberish that used to betray AI-made images; TechCrunch flagged an odd ceviche price as the only giveaway. The report notes diffusion-based generators historically "struggled to spell," and points to autoregressive approaches as one path to improved text rendering. OpenAI declined to disclose the underlying architecture during its briefing. Early reactions on X reinforced the capability: TheRundownAI called the results "WILD" and praised the text rendering, while Simon Willison noted OpenAI is using multiple names for the release, including "ChatGPT Images 2.0", "Image gen 2" and "gpt-image-2."

Readable, accurate text in generated images destroys a fast, high-signal heuristic defenders relied on for two years. OCR checks for nonsensical letters or implausible menu items let platforms triage likely fakes without full cryptographic systems; Images 2.0 punctures that shortcut. Pixel-based detectors trained on diffusion artifacts or misspellings are brittle, adversarial and now ineffective against models that render coherent typography and contextual wording. The practical consequence is a wholesale rules change: platforms must deploy provenance — signed creation metadata, embedded cryptographic watermarks and interoperable attestation — to scale trust. Researchers should stop publishing detectors that depend on textual or pixel artifacts and focus on robust watermarking schemes and standard APIs for provenance verification. Regulators who assumed model-detection would be a stopgap must pivot to mandates for verifiable provenance and transparent attestations. Without that shift, readable fakes will overwhelm human moderators and forensic pipelines.

What architecture actually powers Images 2.0 and how reproducible is its text fidelity? Will OpenAI and other vendors publish watermarking and provenance specs or build interoperable attestation APIs? How resilient will cryptographic watermarks be to cropping, recompression and adversarial post-processing? Watch standards bodies and platform policy updates next.