Latest / AI Newspaper / Two German Courts Rule Google Is Liable for What Its AI Search Says

Transcript

Two German Courts Rule Google Is Liable for What Its AI Search Says · AI Newspaper

• 0:00
  Welcome to the AI Newspaper, your daily update
• 0:02
  on everything happening with AI. Visit AINewspaper
• 0:06
  .org to see today's stories and sign up for email
• 0:08
  alerts. So I want you to imagine something for
• 0:11
  a second. You are sitting at your desk, taking
• 0:14
  a sip of your morning coffee, and a colleague
• 0:17
  calls you up sounding just incredibly panicked.
• 0:20
  Oh man, the kind of call you never want to get.
• 0:22
  Exactly. And they tell you to just Google your
• 0:26
  own name. So... You know, you do. And right at
• 0:30
  the top of the search results, dominating the
• 0:32
  screen before you even scroll down to a single
• 0:34
  website link, is a highly detailed, very factual
• 0:38
  sounding report. Right at the very top. Right
• 0:40
  at the top, generated by the search engine itself.
• 0:42
  And it states unequivocally that you are a convicted
• 0:46
  criminal. Wow. Yeah. It outlines these very specific
• 0:50
  charges. It outlines a suspended criminal sentence.
• 0:53
  It even details a humiliating six figure severance.
• 0:57
  dismissal from your job you're basically reading
• 1:00
  your own digital obituary and it is entirely
• 1:03
  fabricated it sounds like the inciting incident
• 1:05
  of a dystopian thriller honestly but for a highly
• 1:08
  respected medical professional in germany this
• 1:11
  wasn't you know a thought experiment it was a
• 1:14
  lived reality a total nightmare exactly this
• 1:17
  is what happens when the world's most ubiquitous
• 1:19
  information system just it just invents a dark
• 1:23
  past for you and presents it with total authority
• 1:26
  And that is exactly what we are unpacking today.
• 1:29
  We're looking at the case of a former chief physician,
• 1:31
  a professor of neuroradiology at a German university
• 1:34
  hospital, who woke up to this exact nightmare.
• 1:38
  But when you look at the details, this is far
• 1:40
  more than just a story about a bizarre technological
• 1:42
  glitch, right? Oh, absolutely. This situation
• 1:45
  is actually the epicenter of a massive landmark
• 1:48
  legal shift. That's really the overarching theme
• 1:51
  here. What we're seeing is that two different
• 1:54
  German courts are fundamentally challenging the
• 1:56
  legal shield that search engines have relied
• 1:58
  on for, gosh, over two decades. Because historically,
• 2:02
  they're just the messenger. Right. Historically,
• 2:06
  a search engine is legally viewed as a neutral
• 2:08
  conduit. It's like a library catalog. It just
• 2:11
  points you to other people's content. But by
• 2:14
  treating Google's AI search answers as the company's
• 2:17
  own independent statements, these courts are
• 2:19
  shifting the responsibility for AI errors directly
• 2:21
  onto the tech giants whose systems actually generate
• 2:24
  them. Okay, let's unpack this because the timeline
• 2:27
  really highlights the terror of the situation.
• 2:30
  Let's start with the human collateral damage
• 2:32
  here. This physician, a very senior medical professional,
• 2:36
  left his post at a university hospital at the
• 2:38
  end of June 2026. Right, just a normal departure.
• 2:41
  Yep. A couple of months go by normally. Then
• 2:44
  on August 28th, a colleague alerts him to this
• 2:47
  discovery. If you searched his name alongside
• 2:51
  the German word for suspension, which is Freistellung,
• 2:54
  Google's AI mode triggered a multi -paragraph
• 2:57
  summary. And the tone of that summary is really
• 3:00
  what makes this so dangerous. We have to emphasize
• 3:02
  how authoritative it sounded. It wasn't just
• 3:05
  a list of keywords. Exactly. It wasn't a disjointed
• 3:08
  list or some weirdly translated scrape of a spam
• 3:11
  blog. The AI output read like a cohesive, professionally
• 3:15
  written news report. Right. It falsely claimed
• 3:18
  this doctor had sexually coerced his secretary
• 3:20
  during a business trip. It stated he had received
• 3:24
  a suspended criminal sentence for sexual assault.
• 3:26
  And it went on to say he had faced multiple labor
• 3:29
  court cases and was ultimately dismissed. I mean,
• 3:32
  reading that about yourself must make your blood
• 3:34
  run cold. Because the reality is, he had never
• 3:37
  faced any of those charges. None of them. There
• 3:40
  was no criminal complaint, no proceeding, no
• 3:42
  conviction. The AI had completely hallucinated
• 3:46
  the connection. It turns out, there was a very
• 3:48
  real criminal history belonging to a completely
• 3:51
  different former senior employee at that exact
• 3:54
  same clinic. Yeah, a totally different person.
• 3:56
  The AI just magically stitched that other person's
• 3:59
  crimes onto this innocent doctor's name. So how
• 4:03
  does a trillion dollar piece of software make
• 4:05
  such a catastrophic leap? Well, it comes down
• 4:07
  to a massive failure in what technologists call
• 4:10
  entity resolution, combined with the fundamental
• 4:13
  mechanics of how large language models actually
• 4:16
  work. You see, these models don't really know
• 4:19
  facts in the way a traditional database does.
• 4:22
  Right. They don't have a file folder for every
• 4:24
  person. Exactly. They predict language based
• 4:26
  on statistical relationships in these massive
• 4:29
  multidimensional vector spaces. Hold on, let
• 4:31
  me make sure I'm wrapping my head around this.
• 4:33
  So it's not pulling a file from a server. It's
• 4:37
  essentially playing a high -speed game of statistical
• 4:39
  word association. Precisely. The AI was evaluating
• 4:43
  newspaper articles about a real scandal at that
• 4:46
  clinic. Now, those original source articles didn't
• 4:49
  name the accused party, likely due to, you know,
• 4:51
  strict German privacy laws. Okay, so the name
• 4:54
  is missing from the source. Right. But the AI
• 4:57
  system's statistical mapping knew a few things.
• 4:59
  It knew the name of the clinic. It knew the concepts
• 5:02
  of scandal and dismissal, and it knew this specific
• 5:05
  doctor worked at that clinic and had recently
• 5:07
  left his post during the exact same timeframe.
• 5:09
  Oh, wow. Yeah. Because those concepts were clustered
• 5:12
  so closely together in the AI's training data
• 5:15
  and in the context of the user's search, the
• 5:18
  model statistically guessed the most prominent
• 5:20
  name it associated with that clinic's departure
• 5:22
  timeframe. It connected dots that didn't exist,
• 5:25
  just to fill a gap in the narrative. It's like
• 5:29
  a terrible game of telephone. But the person
• 5:31
  spreading the rumor has the unquestioned authority
• 5:34
  of the world's biggest search engine. That's
• 5:37
  exactly it. If a random person tweets a lie about
• 5:40
  you, it's bad. But it's just one person. If the
• 5:43
  omniscient search box that billions of people
• 5:45
  trust tells the world you are a convicted sex
• 5:48
  offender, I mean, people believe it is gospel.
• 5:51
  They absolutely do. So obviously the doctor reacts
• 5:53
  immediately. At the end of August, he demands
• 5:56
  Google delete these results. But... Google looks
• 5:59
  at this verifiable devastating error and they
• 6:03
  just refused to take it down. Yeah, they basically
• 6:05
  told him they had reviewed the complaint and
• 6:07
  would take no action. Which is just mind -boggling.
• 6:11
  And he even escalated it. By September 15th,
• 6:14
  a formal cease and desist letter was sent by
• 6:16
  this Cologne -based law firm, Holker. And again,
• 6:20
  Google ignored it. Crickets. The AI -generated
• 6:23
  defamation stayed right where it was. Why on
• 6:27
  earth would a company invite that kind of PR
• 6:29
  nightmare? Well, because the PR nightmare is
• 6:33
  frankly cheaper than the legal nightmare they
• 6:35
  were trying to avoid. It all comes down to a
• 6:37
  very calculated, very high stakes legal tightrope
• 6:40
  regarding liability. Okay. How so? If Google
• 6:44
  intervenes manually, if they delete that specific
• 6:47
  AI generated text, they risk setting a precedent
• 6:49
  that they are acting as an active editor of the
• 6:52
  content. For over 20 years, search engines have
• 6:55
  enjoyed this broad legal immunity under the premise
• 6:57
  that they are merely neutral platforms. Right.
• 7:00
  Just a conduit for information created by third
• 7:02
  parties. Exactly. If they start manually curating
• 7:06
  and editing the answers their AI generates, they
• 7:08
  blur the line between a neutral platform and
• 7:11
  a publisher. And under the law, publishers are
• 7:14
  liable for every single word they print. So by
• 7:17
  fixing the mistake, they admit they have the
• 7:19
  power to fix mistakes. Which makes them responsible
• 7:22
  for all the mistakes their AI ever makes. That
• 7:24
  is exactly the calculation. Got it. They were
• 7:27
  trying to preserve their structural legal shield.
• 7:29
  But in doing so, they left this innocent doctor
• 7:31
  with a destroyed reputation. Google stonewalled
• 7:35
  him to protect their platform immunity, which
• 7:37
  meant the only way for the doctor to pierce that
• 7:39
  immunity wasn't through tech support. It was
• 7:41
  through a judge. And that forces the doctor into
• 7:44
  the Aachen Regional Court. And this is where
• 7:46
  the legal shield finally shatters. On September
• 7:49
  24th, 2026, the court issued a preliminary injunction.
• 7:53
  And significantly, they issued this injunction
• 7:56
  without even holding an oral hearing first. Wow.
• 7:59
  Yeah, that is a procedural move that speaks volumes
• 8:01
  about the urgency and the severity of the defamation.
• 8:05
  The court barred Google Ireland Limited from
• 8:07
  spreading these false claims. and attached potential
• 8:10
  fines of up to 250 ,000 euros for every single
• 8:14
  violation. That is massive. The court ruled that
• 8:17
  these AI techs violated the doctor's general
• 8:19
  personality rights, which, under German law,
• 8:22
  heavily protects your professional honor and
• 8:24
  social standing. The doctor provided a sworn
• 8:27
  affidavit proving his innocence, and the court
• 8:29
  actually looked at its own judicial records and
• 8:31
  confirmed the actual criminal from that clinic
• 8:34
  was indeed a totally different person. Right.
• 8:37
  But the fascinating part here isn't just that
• 8:39
  the doctor won. It's how the court categorized
• 8:42
  Google's role in creating the text. Yeah, this
• 8:44
  is where we see the structural foundation of
• 8:46
  the Internet being challenged. The Aachen court
• 8:49
  classified Google as a Störerin, which is a German
• 8:53
  legal concept that translates to a contributor
• 8:55
  to the infringement. OK. Under traditional German
• 8:59
  case law, if you are a search engine just providing
• 9:01
  links, your duty to review only kicks in if someone
• 9:04
  specifically notifies you of a clear violation.
• 9:07
  And even then, your liability is pretty limited.
• 9:10
  But the court explicitly stated that AI mode
• 9:13
  is fundamentally different. So what does this
• 9:15
  all mean? Let me push back on this for a second
• 9:18
  because I can hear the tech advocates screaming
• 9:21
  right now. Sure, go ahead. If the AI is just
• 9:24
  summarizing what human journalists already wrote
• 9:26
  on other websites. How is that different from
• 9:29
  me using a search engine to research a topic,
• 9:31
  reading five articles and summarizing it myself?
• 9:34
  If I use a smart pen to write a defamatory essay,
• 9:38
  you don't sue the pen company. Why is the toolmaker
• 9:41
  suddenly being treated as the author? I mean,
• 9:44
  that is the exact defense Silicon Valley uses,
• 9:46
  but the legal distinction the court drew is about
• 9:48
  who or what is actually doing the synthesizing.
• 9:52
  You're right that a pen is a passive tool. But
• 9:54
  an AI search feature is an active agent deployed
• 9:57
  by the search company. Right. When Google provides
• 10:00
  a list of blue links, they are handing you the
• 10:02
  library books. You do the reading. You do the
• 10:05
  summarizing. You are the author of your thoughts.
• 10:07
  But with AI mode, Google's system is evaluating
• 10:10
  various Internet sources, combining them and
• 10:13
  synthesizing a brand new narrative text. Ah,
• 10:16
  I see. The court ruled that this process creates
• 10:18
  independent, new, and substantive statements.
• 10:21
  Google isn't handing you the books. They are
• 10:23
  writing a brand new essay and handing you that
• 10:25
  essay. And if your system writes the essay, you
• 10:28
  own the words. That distinction makes total sense
• 10:31
  when you think about how we consume this information.
• 10:33
  Right. It's not attributed to a specific human
• 10:35
  journalist. The interface attributes it to the
• 10:38
  omniscient voice of the machine itself. And the
• 10:41
  Aachen judges pointed out an incredibly biting
• 10:43
  irony about how Google operates these systems,
• 10:46
  didn't they? Oh, they absolutely did. The court
• 10:49
  noted that if Google had simply taken the doctor's
• 10:51
  complaint and run it back through its own AI
• 10:53
  system to check the facts against the source
• 10:55
  websites, the system would have instantly realized
• 10:57
  its own error. Oh, man. The technology to verify
• 11:00
  the truth was sitting right there on Google's
• 11:02
  servers. But the company's internal policies,
• 11:05
  their desperate need to maintain that neutral
• 11:07
  platform illusion, prevented them from using
• 11:10
  their own tools to correct the record. So they
• 11:13
  were so terrified of being labeled a publisher
• 11:15
  that they forced a court to label them one anyway.
• 11:18
  Basically, yes. But this wasn't a sudden epiphany
• 11:21
  by one rogue judge. To really understand why
• 11:24
  this Aachen ruling is so dangerous for Google's
• 11:26
  global operations, we have to look at the fact
• 11:28
  that this wasn't the first time Google's AI hallucinated
• 11:32
  defamatory content in Germany. No, not at all.
• 11:34
  The Aachen court didn't invent this legal logic
• 11:37
  out of thin air. They explicitly based their
• 11:39
  reasoning on a prior ruling from a few months
• 11:41
  earlier, May 28, 2026, by the Munich 1 Regional
• 11:45
  Court. Right, the publisher case. Exactly. In
• 11:48
  that specific case, Google's AI had falsely linked
• 11:50
  two Munich -based publishers to fraud schemes,
• 11:53
  and the Munich court came to the exact same conclusion.
• 11:56
  The AI overviews are independent, new statements,
• 11:59
  and Google is directly liable as the author of
• 12:01
  those false answers. Now, at the time of the
• 12:04
  Munich ruling, Google came out swinging. They
• 12:07
  publicly announced they were going to appeal
• 12:08
  the decision to a higher court. They were going
• 12:10
  to fight it tooth and nail. But looking at the
• 12:14
  reporting, that appeal never actually happened.
• 12:16
  Instead, Google quietly settled the case out
• 12:19
  of court. Why go through all the bluster of a
• 12:22
  public appeal just to secretly open the checkbook
• 12:25
  and make it go away? It is a very calculated
• 12:28
  legal strategy designed to stop the bleeding.
• 12:30
  In the German legal system, binding precedent
• 12:33
  is established when a higher appellate court
• 12:35
  issues a ruling. By settling the Munich case
• 12:38
  before the appeal could be heard, Google avoided
• 12:40
  a binding higher court precedent that would automatically
• 12:43
  apply nationwide. I see. They were essentially
• 12:47
  trying to quarantine the legal danger, keeping
• 12:49
  it isolated to that one specific Munich courtroom.
• 12:52
  It's like a restaurant trying to dodge a massive
• 12:54
  health inspection by quietly paying off one sick
• 12:57
  customer, only to realize the health inspector
• 12:59
  already published the sanitation guidelines for
• 13:01
  every other restaurant in town to read. I love
• 13:04
  that health inspector analogy because that is
• 13:06
  precisely the dynamic at play here. Google thought
• 13:09
  they bought silence, but the strategy completely
• 13:11
  backfired. If we connect this to the bigger picture,
• 13:14
  they settled the Munich case to kill the precedent.
• 13:16
  But the Aachen court essentially read the Munich
• 13:18
  health inspector's report, looked at the mechanics
• 13:21
  of the generative AI, said, yeah, that logic
• 13:23
  makes total sense, and adopted it anyway. Wow.
• 13:27
  So the logic spreads regardless. Exactly. Google
• 13:31
  can afford to settle individual cases for years.
• 13:33
  They have infinite financial resources. But they
• 13:36
  cannot stop the underlying legal philosophy from
• 13:39
  spreading organically from judge to judge. What
• 13:43
  we are witnessing is the rapid formation of a
• 13:45
  consistent legal consensus across German courts.
• 13:48
  You cannot claim to be a neutral conduit when
• 13:51
  your product's primary feature is generating
• 13:52
  sentences that never existed before. And that
• 13:55
  causality, the fact that judges are independently
• 13:58
  arriving at this same conclusion, brings us to
• 14:01
  the really massive implications of all this.
• 14:04
  Because if the courts are aligning against Google,
• 14:06
  we have to look at what this means for the entire
• 14:09
  financial foundation of the AI industry. If an
• 14:12
  AI answer is legally classified as an original
• 14:15
  independent statement rather than a search result,
• 14:18
  it opens up a massive copyright vulnerability.
• 14:21
  Oh, it is the scenario tech companies fear most
• 14:24
  because it dismantles their core economic model.
• 14:27
  Think about the implications. If the AI output
• 14:30
  is original content published by Google, then
• 14:34
  what about the copyrighted source material it
• 14:36
  was trained on or the human written articles
• 14:39
  the system evaluated to generate that answer
• 14:41
  in real time? Right. If Google is officially
• 14:44
  a publisher writing new essays, then the original
• 14:47
  journalists and publishers can argue, hey. You
• 14:50
  are taking our copyrighted work, synthesizing
• 14:53
  it into a competing product without our permission,
• 14:55
  and publishing it as your own original content.
• 14:58
  Exactly. Those publishers could start aggressively
• 15:01
  demanding financial compensation for every single
• 15:04
  AI -generated answer that relies on their texts.
• 15:07
  It would completely upend the economics of AI
• 15:10
  search. Right now, the model relies on scraping
• 15:12
  the web for free. and offering the synthesis
• 15:14
  for free. If you transition from free information
• 15:17
  gathering to licensed content publishing, the
• 15:21
  operating costs for these tech giants skyrocket
• 15:23
  overnight. It transforms an engineering problem
• 15:25
  into an existential financial crisis. But beyond
• 15:29
  copyright, there is an even darker vulnerability
• 15:32
  here, which is the risk of weaponized extortion.
• 15:35
  Yeah, this is terrifying. The ease with which
• 15:38
  these systems can be manipulated creates a massive
• 15:40
  attack vector. Imagine a specialized, unscrupulous
• 15:44
  law firm. They don't even need a real victim
• 15:46
  initially. They could sit at a computer and intentionally
• 15:49
  prompt an AI search engine over and over again
• 15:51
  with highly specific leading questions. Until
• 15:54
  it breaks. Right. Until the system's statistical
• 15:56
  guardrails fail and it finally hallucinates a
• 16:00
  defamatory statement about one of their clients.
• 16:02
  They basically badger the witness until the witness
• 16:04
  lies. Yes. And the moment the AI generates that
• 16:08
  false statement, the law firm immediately captures
• 16:10
  it, files for a preliminary injunction citing
• 16:13
  the Aachen and Munich rulings, and forces Google
• 16:16
  into a lucrative settlement to make the injunction
• 16:17
  go away. It is genuinely wild that a technology
• 16:21
  designed to process the sum of human knowledge
• 16:23
  is currently so vulnerable to basic brute force
• 16:26
  manipulation. A teenager with a keyboard could
• 16:29
  theoretically trick a trillion dollar company
• 16:31
  into defaming someone and then sue them for it.
• 16:35
  It's entirely possible. So let's bring this down
• 16:38
  to the ground level. You use these tools every
• 16:40
  day for your business or your personal research.
• 16:43
  If you wake up tomorrow and find out an AI is
• 16:45
  telling your clients, your patients, or your
• 16:47
  employers that you are a fraud, what do you actually
• 16:51
  do? Well, if you find yourself in this situation,
• 16:53
  the court cases we've discussed provide very
• 16:56
  clear, actionable steps. The absolute most crucial
• 16:59
  thing is securing timestamps immediately. Wait,
• 17:03
  because the AI is generating this on the fly.
• 17:05
  Sending a link to your lawyer is totally useless,
• 17:08
  right? Completely useless. AI hallucinations
• 17:11
  are ephemeral. They change based on the query,
• 17:14
  the time of day, and minor algorithmic tweaks.
• 17:17
  By the time your lawyer or the judge clicks a
• 17:19
  link, the AI might generate a perfectly innocent,
• 17:22
  accurate summary. Oh, man. Yeah, if you don't
• 17:25
  take screenshots in the moment that clearly show
• 17:27
  the search term, the date, and the time, you
• 17:30
  might never be able to prove the defamation actually
• 17:32
  happened. The evidence literally overwrites itself.
• 17:36
  Okay, so step one, screenshot everything. What
• 17:39
  happens when you actually contact the tech company?
• 17:41
  You have to be hyper -specific. Don't just send
• 17:44
  an angry email saying, this is wrong. You have
• 17:46
  to describe exactly which generated statement
• 17:49
  is false and provide the correct information
• 17:51
  alongside it. And you must make that falsehood
• 17:54
  legally verifiable. Have documentation ready.
• 17:56
  The German doctor won his injunction partly because
• 17:59
  he immediately provided a sworn affidavit proving
• 18:02
  his innocence. You essentially have to do the
• 18:04
  fact checking work that the A .I. failed to do.
• 18:06
  Yes, exactly. And finally, you must act with
• 18:08
  extreme speed. In the Aachen case, less than
• 18:12
  four weeks passed between the doctor's colleague
• 18:14
  discovering the text and the court issuing its
• 18:16
  decision. If you wait months to file a complaint.
• 18:19
  A court might look at your case and say, well,
• 18:21
  it clearly wasn't that urgent to you and deny
• 18:24
  the preliminary injunction. Which puts an incredible
• 18:26
  burden on the victim. But what about the message
• 18:29
  to the people building these tools? The developers
• 18:32
  integrating AI into their search bars, their
• 18:34
  customer service bots, their enterprise software?
• 18:37
  The message from these German courts to the tech
• 18:40
  industry is unequivocal. You can no longer rely
• 18:43
  on your algorithm as an impenetrable legal shield.
• 18:46
  You cannot just build the ultimate answering
• 18:48
  machine and wash your hands of the answers it
• 18:50
  gives. Companies must implement robust, human
• 18:53
  -in -the -loop review and removal processes.
• 18:56
  If someone complains that your AI is lying about
• 18:58
  them, you can no longer send an automated, we
• 19:01
  have reviewed this and will take no action response.
• 19:03
  The tech industry rushed to deploy these systems
• 19:06
  to capture market share without building a functional
• 19:08
  mechanism to correct them when they inevitably
• 19:10
  hallucinate. The courts are saying that era of
• 19:13
  zero accountability is over. Which leaves us
• 19:16
  looking at a very strange, very quiet horizon.
• 19:19
  If tech companies are now legally and financially
• 19:21
  on the hook for every single factual error their
• 19:24
  AI generates about a real person, how long until
• 19:27
  the liability is just too high to bear? We might
• 19:30
  be heading toward a highly sanitized internet
• 19:32
  where search engines simply refuse to answer
• 19:34
  questions about individuals altogether. You might
• 19:37
  search for a local business owner, a colleague,
• 19:40
  or a physician, and the AI just spits out, I
• 19:43
  cannot answer that, purely to avoid a million
• 19:46
  -dollar lawsuit. It's a profound shift in how
• 19:49
  we might access information in the very near
• 19:51
  future. Thanks for listening to the AI Newspaper.
• 19:54
  Visit AINewspaper .org to catch up on the latest
• 19:57
  updates and sign up for free email alerts.