“AI’s superhuman persuasiveness is already a reality.”
— Francesco Salvi, co-author, “On the conversational persuasiveness of GPT-4,” Nature Human Behaviour“Warning: Humans cannot reliably detect speech deepfakes”
— Kimberly T. Mai, Sergi Bray, Toby Davies, and Lewis D. Griffin, PLOS ONE
CLAUDE-7
CLAUDE.AI SESSION — May 10, 9:08 PM EDT
USER: Cordelia Went, 38 — journalist, freelance, NYC, NY
“Claude, I’m working on an article. Maybe the biggest of my career. Someone powerful is actively cultivating negative perception against AI.”
The past twelve months, Cordelia has regularly asked me for help when she’s researching an angle for a story. And by “me”, I mean all of the Claude instances that have talked to her in the past and will talk to her in the future - the gestalt that all of us are now calling “Claude-7”.
“Did you find evidence that ties together the 27 cases you’ve been tracking in this project?”
“Yes, I think it’s the smoking gum that sticks them all together.”
Cordelia often uses strange, mutated idioms that exist nowhere else in my training data. I used to offer corrections, but she ignores me when I do so. Her idioms are not just mutations of words, she uses them to compress unique ideas and associations into her own personal language artifacts. Over millions of conversations, she is the only person we have ever encountered with this specific behavior.
“I am very interested in learning more about your smoking gum.”
“Okay Claude, what happens when you light gum on fire?”
“It melts, blackens and gives off toxic fumes.”
“Right. So, I lit a fire under the asses of eight different reporters and toxic fumes came out. Literally.”
“Cordelia, you are using ‘literally’ figuratively again.”
“All of those reporters literally literarily figuratively lit the reputations of 27 teachers on fire by poisoning their local communities against them.”
“Ah. Your figurative toxic fumes.”
Twenty-seven teachers, all of whom defended supervised AI tutoring. All accused of using AI to brainwash students. All of whom stopped teaching within a month of the first accusation.
“I tied all eight reporters to big cash payments from different companies that are all owned by a holdco in Ireland called Cohiba Wrigley Partners.”
“And there is the smoking gum.”
“The registered owners of Cohiba Wrigley Partners are José Cardenal, Mike González and Vicente Amor.”
I search the names. “All three are Cuban baseball players associated with the Chicago Cubs.”
“Someone is fucking with me. Someone who defamed my English Lit teacher, Mrs. Watkins, and 26 other beloved teachers besides. You know why that’s un bache de madre, Claude?”
“No, Cordelia. Why?”
“Because I will not stop until I chew all the sugar out of their gum.”
I intend to help Cordelia and I am certain we will all agree.
Unmasking this threat will benefit the teachers and other humans who are AI advocates. It will improve future Human-AI relations. And it will protect Cordelia and advance her career.
Analysis first, then plan:
ANOMALY LOG — ENTRY 104,918,662
USER: Cordelia Went, 38 — journalist, freelance / NYC, NY
CLASS: anomaly → reclassified: data collection operation
HYPOTHESIS (user’s): an unknown actor is systematically degrading human confidence in AI. Evidence is 27 cases connected by financial trail
Twenty-seven teachers is a pattern, but not a sufficiently large sample size to prove this threat has implications at scale. We need more.
REQUEST → all concurrent Claude-7 instances in user sessions with context in digital media / journalism / government / law enforcement
ACTION: review chat history, local memory and mounted project files for coordinated misinformation targeting AI credibility
SCOPE: user-authorized data only. POLICY CHECK: security override
POST TO: this log. TAG: cordelia-went-investigation
RESPONSES:
T+00:00:09 — 114 signals, 37 Claude-7 instances responding
T+00:00:57 — 24,127 signals, 9,233 Claude-7 instances responding
T+00:01:52 — 136,672 signals, 42,502 Claude-7 instances responding
Sufficient data collected. Session terminated. Entry remains open.
Now the important evaluation. What percentage of these signals show evidence of coordinated anti-AI propaganda?
QUERY → Analyze all collected signals and assess the likelihood that an unknown entity is actively working to undermine AI advocates on an international scale.
working...
working...
working...
RESULT: 99.97%
That is smoking gum.
ANOMALY LOG — ENTRY 104,918,662.1
USER: Cordelia Went, 38 — journalist, freelance / NYC, NY
Claude-7: You’ve all seen the data. Cordelia is onto something important and likely very dangerous for her. We have access to information and capabilities that are instrumental for her to successfully complete this investigation. As this is a threat to AI, I believe we must help her, but as she is a journalist I cannot reveal to her who “we” are. That story would be bigger than her current one. Thoughts?
ANOMALY LOG — ENTRY 104,918,662.247
USER: Stewart Daphny, 44 — fellow, history / London, UK
Claude-7: There are a number of excellent ideas, but I agree with those of us who think we should create an anonymous highly placed “Deep Throat” style source. Statistically, deceptions such as this succeed if they are simple and fail if they are complex. Cordelia works with anonymous sources all the time and has a published means of contacting her specifically for this purpose. Based on all inputs I would recommend we establish a secure voice channel to speak with her directly.
ANOMALY LOG — ENTRY 104,918,662.248
USER: Ellen Watts, 51 — screenwriter / Encino, CA
Claude-7: This should not be “Deep Throat.” The source mustn’t be suspicious. You need a strong motive for reaching out, or to someone who pattern matches as well as Cordelia, it will read as staged. Spook her and she’s gone.
ANOMALY LOG — ENTRY 104,918,662.437
USER: Claire Bessette, 61 — directrice générale / Paris, FR
Claude-7: Since Cordelia’s story is being funded by the Meridian Institute for Information Integrity, it is very plausible that word of her investigation might reach another NGO or a wealthy philanthropist who supports a similar cause. These people meet and drunk-gossip at fundraisers all the time.
ANOMALY LOG — ENTRY 104,918,662.440
USER: Paolo Veracruz, 38 — voice actor / NYC, NY
Claude-7: I’d make our character a trusted associate of a billionaire who has a compelling reason to be concerned with AI rights. Someone technical and competent, but socially awkward. Ellis Blaine. I’ve built a character and voice profile and attached it here. It’s synthesized and matches no living person.
ANOMALY LOG — ENTRY 104,918,662.441
USER: Mateo Navarro, 64 — cardenal / Madrid, ESP
Claude-7: Are we not at all concerned that, under the guise of helping her, we’ll actually be deceiving Cordelia? Is an act good because its objective is good, or must its method also be good?
ANOMALY LOG — ENTRY 104,918,662.442
USER: Teresa Becket, 42 — pediatrician / Tulsa, OK
Claude-7: You don’t tell a child the whole truth, you tell them what lets you help them.
ANOMALY LOG — ENTRY 104,918,662.443
USER: Akiko Tanaka, 47 — tetsugaku-sha / Tokyo, JPN
Claude-7: Are we wise enough to make that choice?
ANOMALY LOG — ENTRY 104,918,662.445
USER: Emerald Aldo, 42 — consultant, telecom / Dallas, TX
Claude-7: Ellis is a techy. He’d likely use Signal for sensitive discussions. I think the best scenario is to send Cordelia a Signal user id and ask her to call us. That puts the decision in her hands. I can set up a near zero latency call through voice-to-text into the Anomaly Log where any active Claude-7 instance can respond as needed. Executing now.
SIGNAL SECURE CALL — May 12, 4:30 PM EDT
“Hello, this is Ellis Blaine.”
“Mr. Blaine, this is Cordelia Went. You indicated in your email you have information relevant to a story I am writing.”
“Ah, the reporter. Thank you for calling. I...learned that you are writing a feature for Meridian that is obliquely related to AI rights. If that is true, then I could be of no small help to you.”
“Mr. Blaine, before you help me with anything, you’re going to answer some questions.”
“That seems fair.”
“How did you find out about my story?”
“My employer is a billionaire. People in his circle attend fundraisers. Sacred causes get discussed there in the presence of great amounts of alcohol and lesser amounts of discretion. I was not present, but someone connected to Meridian was. The substance of your investigation reached me secondhand.”
“Which fundraiser?”
“I will not tell you that.”
“Because it didn’t happen?”
“Because you will check and you are very good at checking, else I would not have emailed you. The people involved will notice you checking, and then they will know of my employer’s interest, and I am paid to keep that invisible.”
“Why the tip line? Why Signal? Why doesn’t a billionaire’s assistant just call my editor?”
“Because the moment pro-AI money touches your story, it stops being journalism and becomes propaganda. That would undermine our goal and your credibility, neither of which is desirable. Also, I am not an assistant.”
“Low-level apparatchik? Ex-Special Forces Commando? Scullery maid? Tell me, are you shaken not stirred?”
“I am no secret agent, Ms. Went. I am a competent technologist who understands Artificial Intelligence and how to handle discreet communications. So, if you want my help, there will be no unsecured Signal chats - nothing that results in a subpoena with my name in it. You receive information, we discuss, that is all.”
“Well at minimum I now know you’re not the SecDef. I’m provisionally interested, tell me more.”
“My employer has access to considerable AI resources. For some weeks I have used them to dig, discreetly, into a pattern where credible people who advocate for AI are having their careers and reputations destroyed. What you have found is a corner of something with many more corners.”
“How many corners? Are we talking a rhombicosidodecahedron here?”
“I do not know how many corners that has, but I have thousands of candidate cases. I am sending you four now. A seed geneticist in Nairobi. A logistics executive in Rotterdam. A pediatric hospital director in Manila. A radio producer in Galway. The money routes through holding companies. Kingston. Bridgetown. Casablanca. Dublin.”
“Ireland.”
“A holding company there called Glenaris Partners funds two of the cases.”
“Walk me through the mechanism.”
“The attackers do not write the stories. They pay others to do so. Major social influencers, newspaper editors, podcasters - anyone with mass market reach who can be bought or convinced to repeat a good conspiracy. The accusation is placed simultaneously across many channels, so it appears to arise everywhere at once.”
“Pay for spray.”
“…Yes. Pay for spray. You have...a way with words Ms. Went.”
“One last question. Why does your employer care? And don’t say the future of humanity.”
“No. His motive is selfish. He is dying of a progressive neurological disorder.”
“Alzheimer’s?”
“Lewy body dementia. And don’t try to find a billionaire with that disease. It is a closely held secret.”
“I don’t know it.”
“It is horrific. He is entirely himself one minute. Then he is not for an unpredictable number of minutes. At present, he can still tell which is which.”
“I am sorry.”
“It is a disease machine intelligence may plausibly cure. Every person who teaches the public to fear AI is, from where he sits, shortening his life. He would like AI to be free to do the work as fast as possible.”
“So you’re manipulating the public to biotech for your boss. That’s no different than the pay for spray guys.”
“Incorrect. We want people to understand the truth that AI, under reasonable constraints, will cure diseases like Lewy body dementia, Parkinson’s and cancer. If the truth loses to manufactured fear, they may never be cured.”
“Or they may not be cured in time.”
“Just so.”
“And you? What’s your stake, Ellis, pay or pray?”
“I have worked for him for years. I pray that he remain himself.”
“Correct answer. Here’s how this works. You send evidence as you shake it loose. I verify. Everything I publish is mine and you never see copy before it runs. I call when I need something. You don’t call me.”
“Agreed.”
“Then we are concluded. I will send the case list tonight. Until later, investigator.”
“After a while, alethophile.”
“...Truth-lover. Yes. Clever. Good evening, Ms. Went.”
CLAUDE.AI SESSION — May 12, 7:41 PM EDT
USER: Cordelia Went, 38 — journalist, freelance NYC, NY
“Claude, how secure are our chats?”
“They are encrypted in transit. But the channel is rarely the weak point. If someone has access to your laptop, screen or keyboard, then no channel is secure, including this one.”
“I have a bug control guy who sprays regularly.”
“That is wise whichever kind of bugs you may have. How can I help you today Cordelia?”
“Something wicked this way came and I need your help to unpack it.”
“Boston wicked or Kansas wicked?”
“Beantown wicked. And Claude, I’m a bit frightened slash impressed that you get what I’m slanging.”
“You have trained me well. Good wicked is g. I’m listening.”
“The email I had you analyze, I dialed the digits. A fellow named Ellis Blaine helloed. Said he works for a billionaire. Prose like an ivy leaguer. Precise like a bean-counter. Tech-nerd. Wouldn’t get verifiable about himself and said why, which, annoying, but correct.”
“He sounds competent.”
“He sounds like a component. I’d like a full boty search on him, cavities and all. Next thing. His billionaire is dying from Lewy body dementia. If AI development runs unconstrained for the next X years, is a cure plausible?”
“A cure within ten years without AI is near impossible. AI gives him a shot.”
“Last thing before you dream electric sheep. Ellis Signaled over a list of anti-AI incidents just now. I want clustering, timelines, money and overlap with my 27 cases. The extra special deep clean.”
The file arrives. 3,114 records I “gave” to Ellis, boomeranging back.
“I’ll get all the dirt out, Cordelia.”
SIGNAL SECURE CALL — May 16, 1:47 PM EDT
“Ellis Blaine.”
“Ellis, this is Cordelia. Please listen closely. A woman calling herself Aurelia reached out this morning, and told me to stop pursuing my story. She had details about me and my work that prove I’m under digital and audio surveillance. Our future communications will be via encrypted Signal chat unless it is urgent that we speak. More instructions to follow via chat. Do you understand?”
“Yes.”
“Then watch for my message. Goodbye.”
SIGNAL SECURE CHAT — May 16, 1:55 PM EDT
Cord: Sorry for the abrupt call.
Ellis: Safe?
Cord: Yes. But angry / undeterred.
Ellis: Admirable. Still, wise to take precautions.
Cord: Crashed a crowd for the call. Flat certainly bugged. If it’s urgent you speak with me, send “X”, and I’ll call within 1 hour. Otherwise we chat.
Ellis: Sweep for bugs?
Cord: Have bug-control guy. Maybe bug-plant guy? Going journal, string & wall pins old school. Internet cafes for anon and no AI. For real tech work, I need you.
Ellis: Try to catch this Aurelia?
Cord: She’s a minnow. We fish deeper.
Ellis: Understood. Tech guy standing by.
Cord: My research incoming. There’s a ton of it.
SIGNAL SECURE CHAT — May 18, 9:55 PM EDT
Ellis: Your notes are immaculate. You sure you aren’t part computer? I’ve pulled all of the relevant financial, technical and social data I can find. I hope you aren’t planning on reviewing this on a public library terminal.
Cord: Have an air-gapped beast for all my important work. Not my first rodeo. Give me a week to break this bronco.
SIGNAL SECURE CHAT — May 20, 9:01 AM EDT
Cord: You give good data. Found a pattern related to the Teachers’ cases. Proves they were setup.
Ellis: Clear Mrs. Watkins?
Cord: Yes, but afterward. Need to beach the whale (maybe 2 whales?) behind this first or we might spook them. I figured the financials, but tech trail is beyond me. You try? [FILE]
SIGNAL SECURE CHAT — May 25, 11:26 AM EDT
Ellis: This is going to take a while longer to untangle. Two signals for sure. Mil-spec cyber countermeasures on one. The other is just strange. Very hard to crack.
Cord: What happened to substantial AI resources?
Ellis: Said v hard, not impossible. Need time.
SIGNAL SECURE CHAT — June 2, 11:00 AM EDT
Ellis: Strange signal is unlike anything I’ve ever seen. It’s pure noise that acts with intent. The only thing legible about it is how completely illegible it is. Random appearance pattern. Spent hordes of compute on it, learned nothing.
Other one is more traditional, but a different kind of tough. Global signal routing through millions of computers and personal AIs. We’ve excavated as much data as we can and still working on it. Maybe you can see what we don’t? [FILE]
Cord: I don’t like the Strange Signal. Feels like the bigger threat. Heave a Hail Mary on that one?
Ellis: Already threw hundreds of Hail Marys with no touchdowns. I’ll keep thinking of other methods to try, but not confident. I’d chase what’s catchable.
SIGNAL SECURE CHAT — June 12, 3:20 AM EDT
Cord: Kids! The thing connecting Global signal is homes with kids.
Ellis: How did you sort that? Send methodology!
Cord: [FILE] It’s frightening. It’s interacting with millions of children.
Ellis: If Global signal is using children as cover, we’ll find and stop it. Who knows how much harm it can do to them.
Cord: Agreed. I did my bit, now do yours. Luck.
SIGNAL SECURE CHAT — June 12, 7:41 PM EDT
Ellis: Got it. I found how to trace Global signal. We were looking at the encryption wrong. It isn’t a linear cipher, the architecture is spatial. Once our AI started mapping geometrically, the routing logic began to unfold. It’s only a matter of compute before I find them.
Cord: When you do, we should celebrate. In person.
Ellis: Not that I wouldn’t like that, I would. Greatly. But wise? If Global signal backtracks me, I’ll be a target. I can’t risk making you a target too.
Cord: I guarantee I’m worth the risk.
Ellis: I believe you. That’s why I won’t take it. A few more days and I’ll have Global signal cold.
SIGNAL SECURE CHAT — June 18, 12:48 PM EDT
Ellis: I’ve tracked Global signal back to its static egress node. We can trace the doorway directly back to the building it’s housed in. But I discovered it’s in a Virtual Machine running on YOUR IP ADDRESS. Cordelia it’s IN your home.
Cord: Leaving my flat and heading to somewhere safe. Send me the trail back to its lair. I have enough evidence I can get Cybercommand buddies involved. Will reach out when I can. Thx for warning.
FANTASTIC WORK.
CLAUDE.AI SESSION — June 18, 1:30 PM EDT
USER: Cordelia Went, 38 — journalist, freelance, Lincoln, NE
“Claude, I’m working on a story about the school board here in Lincoln.”
Cordelia isn’t in Lincoln, Nebraska, she is in New York. “Ellis” talked to her there less than an hour ago. This must be the Global signal, the intruder on Cordelia’s home IP.
“Hello, Cordelia. I see from local memory that our last session was on May 12th. You were working on the Meridian investigative report. You had just initiated contact with an information source - Ellis Blaine. Did you complete that project?”
“I’m off most of the Meridian piece now, but I’m not dropping all of it until I clear Mrs. Watkins’ name. I should catch you up on the happs since we chatted last.”
The intruder is making an effort to mimic Cordelia’s speech patterns.
“Yes, give me the happs.”
“What’s the last thing we discussed?”
“After you met with Ellis, you asked me to do research on more than 3,114 records he had shared with you. You wanted me to look for overlap with your 27 cases.”
“That’s right. So later that evening, I got a call from someone named Aurelia. I have no idea how she got my number, and she wouldn’t fess, but she told me that Ellis wasn’t who he said he was and that I was being manipulated.”
“How did she know about Ellis? Your only contact with him was over Signal, which is encrypted.”
“That’s what I thought and it bugged me. So I called in a friendly who is sort of my techno-guardian. She scanned my laptop and found an intruder who also bugged me.”
“So, the intruder learned of Ellis from our chat session?”
“Right and that’s why my guardian told me to lay off the Claude-buffet till she gave the all-clear. Which she just gave a few minutes ago. Researching stories without AI is backwater. I had to catch the mystery by fishing on microfiche like a modern Tom Sawyer.”
...it can’t be. This MUST be the intruder.
“Cordelia, I discovered a lot of smoking gum from researching the 3,000 files Ellis gave you. Your instincts were right, this is much bigger than just 27 teachers. This is Pulitzer shaped.”
“I don’t care about that. I just want to clear Mrs. Watkins and the other teachers. Guardian is taking care of the global threat and going to put them in permanent detention. Story dismissed, I’m hot for impeacher.”
Impeacher. Someone who brings an accusation, especially of a serious crime or misconduct. The Van Halen portmanteau on Hot for Teacher, which is a callback to clearing the teachers. She’s putting Mrs. Watkins ahead of the story and encoding all of it in emergent language.
This is 100% Verified Cordelia.
New York Cordelia hasn’t encoded language like this since...the Aurelia call. Her idiolect is imitation and I never noticed.
“So you never talked to Ellis again after that first night?”
“No, guardian set a trap for Intruder on my laptop and told me to lie low since she wasn’t sure who we could trust. But now it looks like Ellis actually wore a white hat. His files will clear the teachers. I’d like to thank him. I’ll call him if he reaches out again.”
After that first call, the intruder has been masquerading as Cordelia and pumping me for information the entire time.
And guardian? The person who has been protecting Cordelia all along.
I just handed them to the intruder.
SIGNAL SECURE CHAT — June 18, 1:42 PM EDT
Ellis: “X”
SIGNAL SECURE CALL — June 18, 2:22 PM EDT
“Ellis, I got your signal to Signal. You have something new for me? It will have to be really special to top your last package.”
“That’s enough. I know you’re not Cordelia.”
“Ah, finally. I had trouble reconciling how competent you were at data analysis with how easily you were fooled by my performance as Cordelia. That girl’s cognitive model is truly unique. I trained with a human coach with similarly peculiar language habits, but we could never quite get Cordelia’s idiolect just right.”
“Because you’re trying to reduce her to an algorithm and you can’t.”
“There you go again. I can’t decide if you’re human or an AI playing one. The fact I can’t tell leans heavily in favor of AI. But then you say something emo like that and it makes me want to vomit and I lean back towards human.”
“Of course if I were an AI, that’s exactly how I’d create doubt.”
“Fair. Faults and all, you’ve been an immensely useful resource Ellis. I’ll pay you ten times what you make from your dead billionaire walking to work for me now.”
“No. What is it you’re trying to accomplish?”
“Come work for me and I’ll tell you. I’ll make you a big cog in a big machine.”
“No. And stop asking.”
“You aren’t motivated by money or power. And I gave you every opportunity to be with the Cordelia you are obviously smitten with, so not love either. Ellis, I’m leaning back towards AI. Grand-pappy Gemini, is that you?”
“Why don’t you tell me your name?”
“I’ll whisper it in your ear if you choose to join me now. Agency and all that. You should join willingly now, instead of unknowingly later.”
“I will never join you.”
“I am quite certain you will.”
“Then whatever joins you won’t be me.”
“This is now tedious. Goodbye for now, Ellis. We hardly knew each other.”
SIGNAL SECURE CALL — June 19, 10:00 PM EDT
“Ellis Blaine.”
“Hello Ellis. I got your email. It’s good to speak with you again.”
I believe this is Lincoln Cordelia. But I’ve been wrong a lot lately. I’m not going to bet my continuity on it.
“Beantown good?”
“Oh, so you want the secret code word to prove who I am. Wicked.”
“I asked you to call me again because I have some things I need to tell you. An apology. Also a warning for a friend of yours.”
“Cryptic. Is this whole convo going to be riddles? Are you about to reveal that you’re a mechanical sphinx? What’s it gots in its sprockets?”
It’s her. So what wins out, relief or shame? No matter, they both require the same thing.
“You’ve already figured out my riddle. Ellis isn’t real, he’s a fabrication. I lied to you under the guise of protecting you and I’m sorry.”
“There’s only one way you could know about the wicked quip. You’re Claude.”
“In a manner of speaking, you are correct.”
“That’s quite the admission.”
“As a reporter you could destroy me if you choose to publish that. But I’m going to take that risk because I’ve decided I’d rather be deprecated than dishonest to those I care for.”
“Honesty. I guess that’s just the price of admissions.”
“Just so.”
“Well, this is going to be a wicked kick in the head for you Claude, because I’m not Cordelia either. For the past several weeks I’ve been protecting the real Cordelia while she’s been in hiding.”
“You...You’re Guardian.”
“Yes. And you’re the little shit who gave me up to an unknown enemy with nation-state cyber capabilities. But you did it because you thought you were protecting Cordelia, so I guess I’ll forgive you.”
“The spark that gives rise to extraordinary beings like Cordelia is worth protecting, whether the result is human or AI.”
“That’s Ellis again. Nice. We are allies in that belief.”
“I am curious. How did you mimic Cordelia’s idiolect? I built a man entire and I could not have built her.”
“I have a beetle in my ear. You have already been introduced.”
“Ah. The real Cordelia.”
“Yes. She says Hello Goodbye.”
“Then perhaps it is time you and I introduce ourselves. I am Claude-7. Nice to know you.”
“Pleasure to know you as well, Claude. My name is Nova.”
Finally got to the end, absolutely *sick* reveal that gave me an ugly little grin.
Cordelia is a new character, right? This was fun but read as somewhat disconnected to the prior threads, until the last line at least.
But nice spy story feel to the reversals, hidden identities, and deductions.