On July 30, 2026, Google added a new generation button to the web
version of Google Earth, powered by its proprietary image
generation model Nano Banana 2. Google’s thinking at the
time was straightforward: they wanted to transform Google Earth from a
database purely for viewing satellite imagery into an interactive tool
for hypothetical scenarios and reasoning. Whether for urban planning,
simulating climate disasters, or geography education, users could simply
navigate to any location on the map (serving as authentic satellite
footage input) and type in a prompt (such as simulating a neighborhood
flooded by heavy rain) to directly synthesize a realistic hypothetical
image on top of the authentic satellite basemap (serving as the
output).
However, just one day after launching the feature, Google urgently pulled it back. In their official statement, Google explicitly stated the reason: “We deeply appreciate the unique trust people place in Google Earth as a reliable window into the real world… but we have also seen people share generated screenshots on social platforms that appear to violate our policies. Before we have stronger safeguards in place, we decided to withdraw this feature… These generated images were not directly presented on anyone else’s main interface, and they also carried an AI-generated watermark.”
Reading this, a natural question might arise: given that Google emphasized in its statement that the images carried AI watermarks, and that these generated results did not directly replace the real maps on anyone else’s screen, why did people still mistake them for events happening in reality?
The key lies in how the images spread after leaving the interface. After users generated hypothetical images on the Google Earth web interface, they would casually take a screenshot and post it to X (Twitter), Reddit, or WeChat groups. Once an image leaves the bounds of Google Earth’s native interface, lightweight AI labels in the corner or invisible watermarks become virtually useless on social platforms: ordinary users scrolling through social media won’t bother using specialized technical tools to scan for watermarks; all they see is a satellite screenshot with Google Earth’s iconic perspective and authentic terrain.
On the night of the launch, open-source intelligence analyst Henk van Ess tried it out and found that safety guardrails could easily be bypassed. In one go, he generated images of non-existent craters in Los Angeles, a flooded US Capitol, a burning Googleplex, a collapsed Eiffel Tower, an Iranian nuclear power plant, and a destroyed hospital in Gaza (according to a report by Ars Technica). When these screenshots were shared on social platforms, audiences defaulted to their past habits, instantly assuming they were real satellite screenshots taken from Google Earth. A remark van Ess made in an interview with the BBC captured the secret behind this phenomenon: “The fake image itself doesn’t need to look seamless; it directly inherits the credibility of the map that gave birth to it.”
This statement points to a detail in interface design that is easy to
overlook. Over the past two decades, journalists, intelligence analysts,
and rescue teams have grown accustomed to using
Google Earth as their baseline source for verifying
reality. When people look at images in this interface, they never doubt
their authenticity—just as you wouldn’t question whether a road exists
when you see it on Google Navigation. When a generation button is
embedded directly into this interface, the map still feels like an
objective record, but its output has become an algorithmic simulation.
Viewing the images out of habit, people mistook synthetic visuals for
factual evidence, causing expectations to become misaligned with the
interface’s actual properties.
A commentary from TechPolicy.Press summed it up vividly: “Google Earth is no longer just a record of reality; it has become a generator capable of manufacturing realistic reality, flattening the boundary between our most trusted reality infrastructure and generative AI.”
The Google Earth incident exposed a pragmatic issue: when generated images circulate on social networks via screenshots, the visual impression alone that they were taken from an authoritative interface is enough to make many people accept fake images as real news. Is it truly feasible, then, to rely on downstream technical means to verify the authenticity of these images?
Once images leave their native interface and get widely forwarded,
tracing their provenance often becomes painstaking. On March 8, 2026,
the Iranian media outlet Tehran Times posted a pair of
before-and-after satellite images on X, claiming that a radar at a US
military base in Bahrain had been destroyed. Fact-checking team BBC
Verify subsequently investigated the images. To the naked eye, the
images actually contained rudimentary flaws—for instance, several small
cars parked outside the base were positioned in the exact same spots
across two satellite images allegedly taken a full year apart.
However, flaws visible to the naked eye were merely suspicious
points; what provided hard technical evidence was the
Google SynthID watermark detector called upon by BBC
Verify. SynthID is a pixel-level invisible watermark developed by
Google. Whenever an image is created using Gemini or Google’s AI tools,
this mark is silently embedded into the underlying pixels, making it
difficult to erase through standard cropping, blurring, or fine-tuning.
The detector confirmed directly from the underlying pixels that this
so-called wartime satellite image was indeed forged or modified using
Google’s AI tools (see BBC Verify
report for details).
This marked the first case of wartime misinformation conclusively proven at a technical level using SynthID, demonstrating that watermark-based provenance is indeed technically viable. The problem, however, is that such tracing relies heavily on someone proactively using specialized tools to check, and it assumes that the watermark signal hasn’t been destroyed while the image was being forwarded across various platforms.
Real-world distribution chains are far from ideal. Take the
industry-promoted C2PA standard, an open-source Content
Credentials protocol spearheaded by companies like Adobe and Microsoft.
Its approach is to embed encrypted digital signatures into image or
video files, recording which AI tool created the asset and what edits
were made. But this approach suffers from a fatal flaw: signatures are
stored in the file’s metadata. As soon as a user takes a screenshot or
uploads the original image to Facebook, Instagram, X, or WhatsApp, the
social platforms strip out all metadata during compression and
transcoding to save bandwidth. This isn’t because platforms
intentionally hide forgers; it’s purely collateral damage from network
distribution transcoding (see SoftwareSeni
technical analysis for details). SynthID, directly
embedded in the pixels, isn’t vulnerable to standard screenshots, but it
is both invisible and bound to Google’s proprietary algorithm—Google’s
detector cannot recognize OpenAI’s watermarks, and vice versa (according
to Ars
Technica hands-on testing). In fact, free removal tools like removesynthid.io and various bypass
tutorials are already available online. Developer Sean Goedecke put
it bluntly: unless everyone adopts signing, C2PA risks becoming mere
window dressing. For instance, out of tens of thousands of images
submitted weekly to the FotoForensics website, often only a dozen or so
contain valid signatures (see Sean
Goedecke’s analysis).
This creates a frustrating imbalance, reflecting the classic saying
that “a lie can travel halfway around the world while the truth is
putting on its shoes”: the sense of trust conferred by major platforms
and authoritative interfaces travels freely with screenshots, whereas
the markers proving an image was generated by AI are lost along the
chain of reposts. The authoritative credibility built up by
Google Earth over twenty years was easily carried away by a
single exported screenshot, while proving it fake requires professional
detection tools and tedious verification workflows.
Tests by BBC Verify confirmed this: SynthID performs
well within Google’s own Gemini and Lens ecosystems, but workarounds
still exist; when encountering images with minor color adjustments or
edits, third-party detection tools frequently fail altogether (see BBC test
results for details). This dilemma can hardly be solved simply by
making watermarking algorithms stronger, because trust propagates
through interpersonal networks far faster than technical verification
signals degrade.
Given that watermarks and technical signals inevitably erode during cross-platform sharing, platforms clearly cannot place all their hopes on ex-post verification. Between 2024 and 2026, major global internet platforms gradually developed three distinct governance approaches when dealing with AI-generated content: outright bans, labeling/downranking, and full embrace. Interestingly, which path a platform chose was far from arbitrary.
Comparing Wikipedia and Snapchat side by
side makes this divergence particularly stark. On March 20, 2026, the
Wikipedia editor community voted overwhelmingly by 44 to 2
to strictly ban using large language models to generate or rewrite
encyclopedia articles, leaving only narrow exceptions for translation
and fixing typos (see Quartz
report and AI
Automation policy analysis for details). The community’s main
concern was that AI would undermine the cardinal rule that every
sentence must have an explicit author and reliable citation, while
risking flooding the web with self-referential AI slop. Meanwhile,
Snapchat took a completely different path: in late 2025,
they rolled out the Imagine Lens filter for free to 350 million daily
active users, generating 38 billion AI filter views, and subsequently
launched a full suite of AI ad generation tools in 2026 (see Snapchat
official announcement).
The root cause of these opposite directions lies in the differing promises the two platforms make to their users. Wikipedia promises a verifiable archive of serious facts where every sentence holds someone accountable; Snapchat sets the expectation of playful AR filters shared between friends. Injecting AI generation into Wikipedia directly destroys its core value proposition; in Snapchat, quirkiness and fun are inherent to the product experience, so users would never mistake spoof filters for reality.
The misstep with Google Earth stemmed from crossing the
exact same line. As a baseline reference tool for geography and mapping,
it carries an immense expectation of factual recording. Embedding a
generation button directly into the main map interface effectively
changed what users were seeing from fact to algorithmic simulation,
while keeping the interface promise unchanged. Had it been designed
instead as a standalone virtual sandbox mode—equipped with prominent
status indicators and side-by-side comparisons with original imagery—the
public reaction would have been nowhere near as severe.
Even within the same tech giant, strategies vary drastically across
different products: Google immediately pulled back on
Google Earth after noticing issues, yet aggressively
promoted generation within the Gemini conversational interface; Meta
strictly polices politically sensitive generated images on Facebook
communities, while encouraging unrestricted creation in its design
software Meta Imagine; LinkedIn enthusiastically launched
an AI article-writing feature earlier, only to abruptly shut it down on
July 31, 2026, replacing it with a report button for suspected AI slop.
What dictates platform policy is not corporate PR values, but the
invisible trust contract between a specific product and its users.
Examining these three paths—banning, labeling, and embracing—through the lens of interface promises reveals that each corresponds to specific real-world dilemmas and governance logics.
In domains where errors carry severe consequences, the most common
defense is a blanket ban. BBC found in internal testing
that AI assistants hallucinated or misstated facts 45% of the time in
news writing, leading them to strictly forbid direct use of generative
tools in news reporting and publishing (see Media Copilot
analysis). Top academic publishers like Elsevier,
Springer Nature, Wiley, Science,
and Cell likewise refuse to recognize AI as paper authors,
with the vast majority disallowing AI-generated experimental images
altogether (see academic
publishing policy summary). Music platforms Bandcamp
and Qobuz share the same stance, refusing pure AI-generated
tracks. What these domains share in common is that once fake content
slips through, the cost of post-hoc retraction and clarification is
prohibitively high.
Relying on AI-generated labels for users to discern truth for
themselves produces awkward results in practice. An experiment by
Toronto Metropolitan University’s The Dais involving 2,472 participants
revealed that subtle labels commonly seen today have almost no impact on
whether users believe or share content; only aggressive full-screen
pop-up warnings proved effective (see The Dais
experiment analysis). An MIT study with over 7,500 participants
similarly found that applying a neutral AI label is far less effective
than explicitly marking content as false, and occasionally missing a
label makes unlabeled content appear more credible (see MIT experiment
report). Platforms are well aware that labeling alone merely
satisfies compliance, which is why most now combine labeling with
financial demonetization and algorithmic suppression: Tidal
labels AI music while stripping royalties to zero (according to a Variety
report); TikTok uses a three-pronged approach of
watermarking, metadata, and user self-disclosure to label over 3 billion
videos (see TikTok
governance data); LinkedIn pairs reporting channels
with algorithmic downranking, blocking hundreds of thousands of
automated bot spam attempts daily (according to a Fortune
report).
Platforms choosing to fully embrace AI have essentially offloaded the
burden of verification onto users or the market.
Adobe Firefly introduced a creative agent (see Adobe
release notes), Canva reached $4 billion in annual
revenue driven by Magic Studio, and Notion made AI a core
selling point—because users employ these tools for productivity and
creativity, never expecting their output to serve as authoritative
facts. Epic Games Store refuses to mandate labeling for AI
game assets, with CEO Tim Sweeney criticizing Steam’s mandatory
developer disclosures as overly harsh (according to a VGC
report). Yet public data from Steam shows that games with AI labels
receive 53% fewer reviews on average, indicating that players do care
deeply—omitting labels simply offloads the risk onto buyers.
The root cause of this widespread headache is that AI generates
content far too quickly for human moderation to keep pace. Starling Lab
calculated that while humanity took 149 years to produce 1.5 billion
photographs, generative AI created the same volume of images in just 18
months (see Starling
Lab data analysis). LinkedIn blocks more automated spam
in a single day than Facebook banned accounts over an entire month
(according to a MediaPost
report). Although half of Gen Z users report wanting to unfollow
accounts upon seeing AI slop, social platforms remain flooded with
thousands of posts sustained by automated content farms (see Value
Add VC report). Web2’s legacy moderation model of banning individual
accounts collapses when confronted with identity-decoupled mass
generation.
Regulatory pressure is mounting simultaneously. On August 2, 2026, EU AI Act Article 50 and the California SB 942 regulations took effect on the exact same day. Europe mandates that synthetic content carry machine-readable markers, backed by fines of up to €15 million or 3% of global revenue; California requires platforms with over 1 million monthly active users to provide free verification tools and implicit watermarks. Google Earth’s emergency feature shutdown occurred just two days before these regulations took effect.
Looking back at our own products to evaluate AI governance
strategies, three metrics prove particularly useful: whether the
interface leans toward objective, factual record-keeping; the severity
of consequences if users misinterpret generated visuals; and how easily
generation markers get stripped during dissemination. Had this framework
been applied when designing Google Earth, the safest
approach would clearly not have been embedding a button into the primary
reference interface, but rather creating a standalone virtual sandbox
mode complete with clear status indicators and side-by-side comparisons
with real satellite imagery. As long as the interface’s original promise
remains unviolated, users won’t form mistaken expectations.
To be sure, many industry peers view platform governance purely as calculated self-interest, dismissing interface promises as post-hoc justifications. For instance, Snapchat rolled out features to boost dwell time, Tidal cut royalties to protect musicians’ earnings, and LinkedIn purged slop to prevent its professional network from losing value. While commercial incentives explain why platforms act when they do, economic motives alone fail to account for why different products under the same tech giant adopt polar-opposite stances. In reality, these two perspectives are not contradictory; they complement each other.
Ultimately, obsessing over whether an image or paragraph was generated by AI rarely yields a satisfying answer. What truly determines product safety and governance strategy is how generative technology alters content attributes, what implicit trust the interface conveys to users, and what actions people will take based on that content before uncovering the truth. When product builders stop relying on silver-bullet watermarks or labels and instead re-examine the trust contract between their interface and their users, the challenge of AI content governance finally gains a practical solution.