Can we detect deepfakes reliably yet?
No, not reliably. Detectors work reasonably on known generators and degrade badly on new ones, on compressed social media re-uploads and on partially edited media. Content provenance signing is the more promising route, and adoption is still partial.
Last reviewed 1 August 2026 by the FutureReady editors.
Technology Readiness
Genuinely usable, but with caveats, supervision or narrow conditions.
Readiness verdict
Can we detect deepfakes reliably yet?
Genuinely usable, but with caveats, supervision or narrow conditions.
Screenshot-friendly summary — quote it anywhere, ideally with a link back.
What's possible today
- Detection tools with decent accuracy on in-distribution content
- Cryptographic content credentials attached at capture or export by some cameras and tools
- Platform labelling of declared AI content
What still can't be done reliably
- Accurate detection of unseen generators
- Surviving re-compression, cropping and screen recording
- Confident verdicts suitable for evidence or journalism
Readiness breakdown
Weighted dimensions behind the 54/100 score. These are indicative editorial assessments, not measurements.
- Technical capability55 · weight 25%
Does the underlying technology fundamentally work, even in ideal conditions?
- Reliability35 · weight 20%
Does it work repeatably in messy real-world conditions, without babysitting?
- Accessibility60 · weight 15%
Can an ordinary person actually obtain it and use it without specialist skill?
- Affordability70 · weight 15%
Is the total cost realistic for the person or business who wants the outcome?
- Autonomy50 · weight 10%
How much of the task completes end-to-end without human intervention?
- Regulatory & practical availability55 · weight 15%
Is it legal, supported and purchasable in most major markets today?
Privacy & risk considerations
- False accusations from over-trusting detector scores are a serious harm
- Absence of a credential does not prove fakery
What would need to happen next
- 01Broad adoption of provenance standards across cameras and platforms
- 02Detection benchmarks that generalise
Example products & services
Content provenance standards
Industry consortium
Signed capture-to-publish metadata, partial adoption.
Listed as representative examples, not endorsements. See our affiliate and sponsorship disclosure.
Frequently asked
- Can I trust a detector's percentage?
- Treat it as weak evidence. Corroborate with provenance data and original-source verification.
Related capabilities
Have AI answer my phone calls
Can AI answer my phone calls yet?
Yes, mostly. Consumer phones and cheap call-screening services will pick up, transcribe and hold a short conversation with a caller, and they handle spam and simple appointment calls well. Anything nuanced — an emotional caller, a complex booking, a call that needs your judgement — still gets handed back to you.
Translate a live conversation
Can technology translate a conversation in real time yet?
Yes, and it is genuinely good. Phones, earbuds and some glasses translate spoken conversation with a second or two of lag across major language pairs. It is excellent for travel and everyday chat, and still risky for legal, medical or contractual conversations.
Have an AI do my weekly food shop
Can AI do my weekly food shop yet?
Partly. An assistant can plan meals, build a basket and reorder your usual items, and some retailers offer their own AI basket builders. Fully hands-off shopping — choosing substitutions the way you would, spotting a bad price, adapting to what is actually in stock — still needs you to check the basket before checkout.