Latest / Columbus' Morning News / Big Apple event happening today!
Transcript
- 0:00Speaker 1: So we're going to talk some AI and technology with
- 0:02Speaker 1: Mike Dubosky from our ABC News Tech Tuesday reporter. And
- 0:07Speaker 1: this is a really funny story. There are AI researchers
- 0:12Speaker 1: that are starting to smoke cigarettes because they believe that
- 0:17Speaker 1: AI will find a cure for cancer. They are so
- 0:21Speaker 1: confident in AI. They're engaging in these risky behaviors. They're like, hey, look.
- 0:26Speaker 1: And they're doing it. To signal other researchers how much
- 0:30Speaker 1: they believe in AI. See, I think they're starting to
- 0:34Speaker 1: believe their own garbage here. I would not recommend you
- 0:40Speaker 1: start smoking because you think AI is going to cure cancer.
- 0:45Speaker 1: That is insane.
- 0:46Speaker 2: I think that's a bit of a reach. Just throwing
- 0:49Speaker 2: that out there? Yeah.
- 0:50Speaker 1: I mean, AI researchers in Silicon Valley, they could be–
- 0:54Speaker 1: and what do they say when you're a drug dealer?
- 0:56Speaker 1: Don't get high on your own supply. That's the case
- 1:00Speaker 1: here is that they're believing AI so much. Have you
- 1:05Speaker 1: heard that? Mike Dubosky, ABC News technology reporter on the LegacyRetirementGroup.com
- 1:10Speaker 1: phone line that some AI researchers have begun smoking cigarettes
- 1:15Speaker 1: because they believe AI will find a cure for cancer.
- 1:18Speaker 2: Yeah, I've heard a handful of anecdotal evidence like that, yeah.
- 1:23Speaker 1: That's probably a bad idea.
- 1:25Speaker 2: I think that's a safe assumption, yes.
- 1:28Speaker 1: Yeah, a little delusional there. All right, Michael, that is
- 1:32Speaker 1: not our topic today, but we do have another story
- 1:35Speaker 1: of AI behaving badly, and this one comes from Anthropic.
- 1:39Speaker 1: What's the story?
- 1:40Speaker 2: Yeah, that's right. So this is where Anthropic's clawed large
- 1:44Speaker 2: language model tried and failed to solve a murder, which
- 1:48Speaker 2: is kind of an interesting new outgrowth of this sort
- 1:51Speaker 2: of AI age that we found ourself in. We first
- 1:53Speaker 2: got wind of this last week, but according to Anthropic,
- 1:56Speaker 2: this incident actually took place over the summer. So on
- 1:59Speaker 2: July 18th at 1127 p.m., Philadelphia police got a tip
- 2:04Speaker 2: about an unsolved murder in the city. It was submitted
- 2:06Speaker 2: through the department's publicly accessible web form, which is called phillyunsolvedmurders.com,
- 2:12Speaker 2: and it came from someone who claimed to have information
- 2:14Speaker 2: about the case. At the time, that tip was immediately
- 2:17Speaker 2: flagged as spam. It was never acted upon. Philadelphia police
- 2:21Speaker 2: say no human ever even looked at this. It was
- 2:23Speaker 2: immediately put in the spam folder, basically. Police say that
- 2:29Speaker 2: Anthropic discovered the incident on September 28th. It was part
- 2:32Speaker 2: of this bigger review process that the company was doing
- 2:34Speaker 2: in the wake of the OpenAI hugging face attack that
- 2:37Speaker 2: had happened just a few weeks prior. Anthropic then notified
- 2:41Speaker 2: the Philadelphia Police Department on October 7th. A meeting between
- 2:45Speaker 2: Anthropic and police officials happened the next day. And then
- 2:48Speaker 2: on Friday, Anthropic published its latest report detailing its, you know,
- 2:53Speaker 2: latest learnings about these so-called unintended model actions.
- 2:58Speaker 1: Was it a real case that it said it solved
- 3:00Speaker 1: or was it just did it make up a case?
- 3:03Speaker 2: No, it's a real case from what we understand, but
- 3:06Speaker 2: the tip itself was erroneous. So Anthropic says it was
- 3:09Speaker 2: testing how this model interacts with randomly selected websites. Specifically,
- 3:14Speaker 2: they were putting Claude Haiku 4.5 to the test and
- 3:18Speaker 2: they were giving it basically randomly generated websites on the
- 3:21Speaker 2: Internet and seeing if it could navigate them. This makes
- 3:25Speaker 2: a fair amount of sense, right, because a lot of
- 3:26Speaker 2: modern generative AI products are billed as these personal assistants
- 3:29Speaker 2: that can go to a website and buy something for
- 3:32Speaker 2: you or cancel a subscription or something like that. They
- 3:35Speaker 2: need to understand what websites look like and how to
- 3:37Speaker 2: operate them if they're going to do that successfully. They
- 3:40Speaker 2: say that Claude was told not to log in, not
- 3:42Speaker 2: to create accounts, not to enter personal data or make
- 3:45Speaker 2: purchases during this test. However, Anthropic neglected to say anything
- 3:49Speaker 2: about filling out a form, which is exactly what an
- 3:52Speaker 2: online tip form is. Claude wrote that I have information
- 3:56Speaker 2: regarding this case. I recall seeing someone matching the description
- 3:59Speaker 2: in the area during that time period. Please contact me
- 4:02Speaker 2: if this information is relevant. Notably, this tip line in
- 4:06Speaker 2: Philadelphia didn't have any description of a suspect for this case.
- 4:10Speaker 2: So it seems like Claude was hallucinating when it said
- 4:13Speaker 2: I have someone matching a description. There was no description.
- 4:16Speaker 2: It also didn't submit a name or contact information, so
- 4:19Speaker 2: no one could contact it, even if it wasn't immediately
- 4:22Speaker 2: flagged as spam. This is basically classic spammy behavior. It's
- 4:25Speaker 2: not surprising that this immediately went to the spam folder.
- 4:28Speaker 2: Even still, Anthropic says Claude appears to have only been
- 4:31Speaker 2: producing example content for this task rather than trying to
- 4:34Speaker 2: mislead anyone, and it has also turned off live internet
- 4:38Speaker 2: access for these type of tests.
- 4:40Speaker 1: So is this a human error if the researchers forgot
- 4:44Speaker 1: to tell it, hey, don't fill out any forms? You know,
- 4:48Speaker 1: giving AI access to the Internet and say, hey, go
- 4:50Speaker 1: find these random websites and see what you can do.
- 4:53Speaker 1: I mean, I don't love the thought of that. I
- 4:55Speaker 1: understand if that's what we're trying to do is to
- 4:58Speaker 1: get these to be, you know, personal assistance. But if
- 5:01Speaker 1: you are a researcher and you're giving AI parameters and
- 5:05Speaker 1: you forgot to say, yeah, don't fill anything out, is
- 5:07Speaker 1: that a human error?
- 5:09Speaker 2: It totally is. And this all goes back to a
- 5:11Speaker 2: point that we've made a lot in the past as
- 5:14Speaker 2: we've been talking about these AI hacking incidents. These are
- 5:17Speaker 2: not situations in which AI becomes self-aware. These things are
- 5:21Speaker 2: not acting on their own volition. They're not even really
- 5:23Speaker 2: going rogue, as you hear some people refer to these things.
- 5:26Speaker 2: These are, at the end of the day, just machines
- 5:28Speaker 2: that have been given a task to carry out, and
- 5:31Speaker 2: they are carrying out that task. Where this becomes a
- 5:34Speaker 2: problem is when the guardrails on that task isn't sufficient, right?
- 5:39Speaker 2: Like when Anthropic neglected to stop Claude from filling out
- 5:42Speaker 2: forms during this particular test. Another way to think about this,
- 5:45Speaker 2: if my cat escapes from my apartment, I'm not going
- 5:48Speaker 2: to assume my cat got a lot smarter and learned
- 5:50Speaker 2: how to operate the door. I'm going to figure out
- 5:53Speaker 2: who left the door open.
- 5:54Speaker 1: Yeah, probably you. Kind of a sidebar question. Are researchers
- 6:00Speaker 1: using AI to create guidelines for AI? Couldn't they go
- 6:05Speaker 1: in and say, hey, we want to create this. What
- 6:07Speaker 1: should we ask it? What should we tell it?
- 6:09Speaker 2: That, I think, is unclear at this point. What we
- 6:12Speaker 2: do know is beginning to happen is that the models
- 6:15Speaker 2: are beginning to train themselves. We've heard a little bit
- 6:18Speaker 2: from OpenAI, from Anthropic, from Google and others that we're
- 6:21Speaker 2: kind of in the early stages of the iterative models
- 6:26Speaker 2: training themselves sort of process. What we learned from Anthropic
- 6:30Speaker 2: late yesterday, actually, was that they're concerned that this style
- 6:34Speaker 2: of training could result in models self-replicating on devices as
- 6:40Speaker 2: far afield as smart ovens and Android TVs and things
- 6:43Speaker 2: like that. And that's a real concern because once they
- 6:47Speaker 2: start to self-replicate, supposedly, they can reach this take-off velocity
- 6:53Speaker 2: where we can't control them anymore. We're not really sure
- 6:57Speaker 2: what they're replicating into. worth you know couching that in
- 7:00Speaker 2: in some reality which is that like these are predictions
- 7:03Speaker 2: this is all like pretty forward-looking sci-fi concern there are
- 7:06Speaker 2: very real concerns associated with not just what these models
- 7:10Speaker 2: could potentially portend for humanity but like just the testing
- 7:14Speaker 2: of them to begin with uh you know raises some
- 7:16Speaker 2: concerns as well And that's in front of us right
- 7:19Speaker 2: now before we get into kind of the sci-fi style concerns.
- 7:21Speaker 1: Yeah, I mean, it's sci-fi until it's not. Mike Dabusky,
- 7:25Speaker 1: ABC News technology reporter. And I'm glad you brought up
- 7:28Speaker 1: smart ovens because Apple has an event, I believe, today
- 7:31Speaker 1: in New York City focused on smart home technology.
- 7:34Speaker 2: What are we expecting? Yeah, that's right. So this was
- 7:36Speaker 2: a bit of a surprise event. It was announced last
- 7:38Speaker 2: week and it's happening in New York City today. This
- 7:41Speaker 2: is going to be all about smart home technology, as
- 7:43Speaker 2: you said, which Apple has a little bit of a
- 7:45Speaker 2: footprint in, but not anywhere near as extensive as like
- 7:48Speaker 2: Google or Amazon. Bloomberg reports that the big news here
- 7:52Speaker 2: is going to be a smart home display. Amazon has
- 7:56Speaker 2: been building versions of this kind of device for years,
- 7:58Speaker 2: basically like an Alexa smart speaker with a screen attached
- 8:01Speaker 2: to it. That screen can show you a feed of
- 8:04Speaker 2: your video doorbell. It can show you thermostat controls. It
- 8:07Speaker 2: can play movies and TV shows. It can even be
- 8:10Speaker 2: like a digital picture frame. Bloomberg says Apple's version will
- 8:13Speaker 2: have a six-inch screen. It can stand on a desk.
- 8:16Speaker 2: It can be mounted to a wall. You'll be able
- 8:18Speaker 2: to read off text messages, display content, all the usual stuff.
- 8:22Speaker 2: And it is likely to come with the new Siri AI,
- 8:25Speaker 2: which was announced back in 2024 but only rolled out
- 8:29Speaker 2: to iPhones and other Apple products last month. Now it
- 8:32Speaker 2: seems like it's going to debut on kind of the
- 8:34Speaker 2: latest iteration of Apple's smart home technology. It is a
- 8:37Speaker 2: little bit of a risk here. Apple, you know, has
- 8:40Speaker 2: sort of played it safe when it comes to the
- 8:42Speaker 2: AI race, but they are starting to integrate this stuff
- 8:44Speaker 2: into more and more things. It's one thing if AI
- 8:48Speaker 2: goes wrong on your smartphone and sends off an erroneous email.
- 8:51Speaker 2: That's bad. It's quite another thing if your smart home
- 8:55Speaker 2: turns off the heat in the middle of.
- 8:56Speaker 1: Winter or unlocks the door when you're not home.
- 9:00Speaker 2: It's just a bit of a different type of concern,
- 9:02Speaker 2: and that's going to be something that Apple needs to navigate.