Many AI researchers think fakes will become undetectable
Advances in generative AI are making images, video, and text increasingly hard to distinguish from authentic content, raising fears that deepfakes will soon evade even specialized detectors. Commenters debate technical countermeasures such as AI-based detectors, hardware signing of camera output, and provenance systems like the Content Authenticity Initiative, while noting these can be subverted and may create a false sense of security. Many argue the deeper issue is societal: as digital media becomes broadly untrustworthy, trust will shift toward verified sources, legal and regulatory frameworks, and social norms around consent, identity, and evidence.
Detection vs. Undetectability
- Some argue models can be trained to reliably detect outputs of specific generators (e.g., “detector per model”), since AI is good at spotting subtle artifacts.
- Others note the adversarial nature: generators will adapt to detectors, leading to an arms race similar to GANs or malware/antivirus.
- Targeted detection is seen as more feasible than general “detect any AI image” tools, but critics say you can never be sure you’ve covered all generators.
- Several commenters expect that, eventually, detection will be unreliable and high‑quality fakes will “win” for many real‑world users.
Digital Signatures & Provenance
- Many see cryptographic signatures and provenance tracking as essential: signing at the camera or software level, with some form of decentralized PKI or web-of-trust.
- Hardware keys in cameras, blockchain hashes, and initiatives like content authenticity standards are cited as emerging tools.
- Skeptics stress: secrets can be extracted, hardware/DRM can be subverted, and such systems may create a dangerous false sense of security.
- Some emphasize that signatures only prove origin/control of keys, not truthfulness or absence of manipulation.
Trust, Society, and Institutions
- Broad concern that undetectable fakes will erode trust in digital media, increase fraud, and overwhelm average users, even if experts can still detect some fakes.
- Others say images have never been fully “objective,” and we should shift focus from content to source credibility, as with text.
- There is disagreement on whether established media and institutions can remain trustworthy arbiters; some see them as part of the problem.
- A minority claim this will be a “non‑problem” in practice, akin to existing impersonation and propaganda, while others see societal risk as “off the charts.”
Law, Ethics, and Harm
- Deepfake abuse, especially sexual deepfakes and those involving minors, is highlighted as already harmful and likely to worsen.
- Some propose strong regulation: mandatory AI labeling, making impersonation illegal, and severe penalties for fakes, especially by powerful actors.
- Others question enforceability, given open models and state-level capabilities.
Adaptation & Future Norms
- Expectation that people will eventually treat all audio/video like text: inherently untrusted without provenance.
- Concerns about a “danger window” before that norm sets in, when realistic deepfakes can cause maximum damage (e.g., blackmail, incitement, legal misuse).