Lying becomes the norm
In our article on the great failures of AI, we mentioned that the Rabbit R1 relied on a CES demo that was largely faked. But Rabbit is not an isolated case. Since the explosion of generative AI in 2023, faked demos have become a systematic marketing strategy, used by the biggest players in the sector. And the general public keeps falling for it.
Here are the five most emblematic cases, and the signals that would have made it possible to spot them in advance.
Case 1: Rabbit R1 and the false magic of the "Large Action Model"
At CES 2024 (the Consumer Electronics Show, the world's largest annual consumer electronics trade fair, held every January in Las Vegas), Rabbit unveiled its R1, a small orange $199 gadget that promised to revolutionise mobile computing. On stage, founder Jesse Lyu showed how the device orders an Uber, books a restaurant, buys concert tickets, all by simple voice command. The audience was won over, 100,000 units sold in a few weeks.
On delivery, the reality was quite different. The device could do almost none of what had been shown. The demos were largely pre-recorded scripts (instructions coded in advance, which simulate intelligence but only react to predetermined sequences). The mysterious "Large Action Model" or LAM (a type of AI supposedly able to understand intentions and act on interfaces) that Lyu boasted about was in reality just a wrapper (a software shell that calls other existing models) over standard models with a few hardcoded scripts (coded in hard, not modifiable by the AI).
When Rabbit refused to make its SDK (Software Development Kit, a set of tools allowing third-party developers to build on the technology) available at launch, claiming it was "protecting the user experience", that was a red flag. A genuinely new technology invites developers to extend it. A fake technology hides its code. The rule holds for all AI products: if the company doesn't let developers see what's underneath, there's probably not much underneath.
Case 2: Google Gemini and the "miraculous" video
December 2023, Google publishes a demonstration video titled "Hands-on with Gemini" showing the AI responding in real time to visual and voice questions. The result is impressive: Gemini recognises a hand playing rock-paper-scissors, follows a moving duck, understands complex gestures. The video goes viral, passing 2 million views in a few days.
A few days later, Google admitted it: the video had been heavily edited. The questions had been typed as text, not asked out loud. The answers had been generated from static images, not from real-time video. The actual latency times (the delay between question and answer) had been cut. Bharat Mediratta, then vice-president of engineering at Google, confirmed on X (formerly Twitter) that the video was "a highlight reel" and not a demonstration under real conditions.
The affair caused a scandal in the tech world, but barely reached the general public. Most people saw the original video and came away with the impression that Gemini was far more advanced than it really was.
Case 3: Meta Ray-Ban Display and the crash in front of Zuckerberg
September 2025, Meta Connect conference (Meta's annual event dedicated to new products). Mark Zuckerberg, CEO and founder of Meta (the parent company of Facebook, Instagram and WhatsApp), is on stage to present the new Ray-Ban Display, the future of smart glasses. A content creator is cooking, and Meta's AI is supposed to guide her in real time. Except nothing works. The assistant does not respond, or responds off-topic, or skips steps. The demo collapses live.
Worse, the next segment shows the Neural Band, a wristband supposed to control the glasses through neural sensing (by reading the electrical signals from the muscles in the wrist). The alarm keeps ringing despite Zuckerberg's attempts to turn it off. The room, silent, watches the CEO of the tech giant flounder publicly.
Meta later explained that the bugs were due to a "race condition" (a common programming problem in which several processes access the same resource simultaneously and produce unpredictable results), not to a fundamental flaw in the product. Andrew Bosworth, Meta's CTO (Chief Technology Officer), clarified on Instagram that these were "demo failures and not product failures". The distinction is defensible. But it leaves an open question: if Meta cannot get its own products to work under the controlled conditions of a keynote, what will happen with the end user?
Case 4: Sora and the promise of 60 seconds in 4K
February 2024, OpenAI unveils Sora with a series of stunning videos: a mammoth walking in the snow, Tokyo streets filmed at sunset, an astronaut in a Martian landscape. The videos last up to 60 seconds, in 4K (ultra high definition resolution, around 4,000 pixels wide). Sam Altman, CEO and co-founder of OpenAI (creator of ChatGPT), speaks of a "new era of video creation".
When Sora finally launched, users noticed the difference. The model generated videos of far lower quality than the promotional examples. Durations were capped at 10 seconds for most accounts. 4K was only available to Pro subscribers. And the official demos, as several former OpenAI engineers on Sora acknowledged, were the result of enormous selection (in English, this is called cherry-picking: choosing the most flattering examples while hiding the many failed attempts). Thousands of generations were produced, and only the best 10 or 20 were shown to the public.
As we documented in our article on the great AI failures, Sora was eventually shut down in April 2026, with total revenue 7 times lower than the cost of a single day. The marketing promise had never met economic reality.
Case 5: Apple Intelligence and the Siri that did not exist
The Apple case, already covered in our article on the great AI failures, deserves a specific mention here. At WWDC (Worldwide Developers Conference, Apple's annual developer conference) in June 2024, Apple showed on stage a new Siri capable of using personal context, acting across apps, and understanding complex situations. Every iPhone 16 was sold from September 2024 with these features announced.
These features did not exist. Apple had to admit this publicly in March 2025, pushing the new Siri back to 2026. The class action (a procedure in which many claimants group together to bring a joint lawsuit) filed by iPhone buyers was settled in May 2026 for 250 million dollars. The main accusation was that Apple had marketed "AI capabilities that did not exist at the time, still do not exist, and will not exist for two years or more".
This is probably the most serious case on this list, because it is not a misleading demo at an event, but a global marketing campaign spread over months, with advertising across every medium, announcing features that were entirely fictitious at the time of sale.
The five signals to recognise
If you look at these five cases as a whole, several warning signs recur. Here is how to spot them in advance.
Signal 1: no testing period for the independent press. The Rabbit, Sora and Apple Intelligence demos were never tested in real conditions by independent journalists before launch. When an AI product is released without a public testing phase, it is often because the real result would disappoint. Real technology welcomes independent testing.
Signal 2: the demo is too smooth. Real AI in 2026 still has noticeable latency, hesitations, limits. When a demo shows instant, perfect answers with no approximation whatsoever, there is a 90% chance it is editing or a script. Credible demos show, on the contrary, the moments when the AI hesitates or corrects itself.
Signal 3: the exact conditions are not specified. The Sora and Gemini demos did not say how many attempts had been needed to get the result shown. The Rabbit demos did not specify whether the actions were pre-scripted. When a demo refuses to detail the exact process, it is often because the process would not survive transparency.
Signal 4: the source code or the API is closed. The API (Application Programming Interface) is the technical interface that lets other software use a service. As mentioned above, Rabbit refused to give access to its SDK. Apple refused for 20 months to open the new Siri to third-party developers, "because it wasn't ready". A product that really works lets developers extend it quickly. A product that doesn't really work barricades itself.
Signal 5: the promises exceed what the competition can do. When a player announces performance 10 times better than what the world's best labs can achieve, there are two possibilities: either a major scientific breakthrough (rare, and generally accompanied by scientific publications), or a manipulated demo. Sora 60 seconds in 4K in 2024 went far beyond what Chinese and American labs were able to produce at the time. The gap was a signal.
How to navigate this landscape
For a user, a corporate buyer or an investor, the practical rule is simple: never believe an AI demo before it has been tested by at least three journalists or independent creators in real conditions. Wait for the first public tests. Read the comments from users who received the product. Check whether the product really does what was promised on stage.
For developers and entrepreneurs, the lesson is different. As we showed in our article on the AI startups that are winning, the companies that really succeed in 2026 are those that deliver what they promise. Cursor (AI code editor), Perplexity (conversational search engine), ElevenLabs (voice generation), Anthropic (creator of Claude): none of these successes was built on a rigged demo. All of them let their product spread by word of mouth from professionals who tested it themselves.
This analysis complements our series on the blind spots of AI in 2026. To continue, you can read our article on the AI freemium trap, and more broadly our analysis of the AI bubble, which places these demos in a wider economic context.
AI in 2026 does not lack real technologies. It lacks transparency about what really works. Rigged demos artificially prolong the confusion. Learning to spot them means gaining two to three years' head start on naive buyers who will discover the disappointments when the bill arrives.