AI ALERT: Genealogists need to be aware of AI shrinkflation and subsidies

ChatGPT's image of the "rug pull".

Picture this: you're working away using an AI agent on a kinship analysis and are suddenly cut-off from finishing.You had checked your usage levels before you started the project and had 100% usage available. "I couldn't have tokenmaxxed, what happened?" You've just had your usage throttled by an AI company through no fault of your own - but theirs. This occurred last weekend with Grok Bot. Some customers were sleeping and hadn't even used it and were throttled. When a customer posted on social media asking for a reset, Elon Musk replied to the post,“The cost of serving Grok Bot was unsustainably high relative to revenue." Grok Bot is included from the $30 Premium tier of the X Subscription upward, with about 800,000 paid weekly users. As reported by TechRadar, one Reddit thread dubbed Musk’s throttle an “indirect rug pull”, but a more common term also applies "Shrinkflation". Consumers are familiar with shrinkflation even not ever hearing this term: it's when a Taco Bell taco shrinks in size but stays the same price, or toilet paper stays the same price but goes from 425 sheets per roll to 320 sheets per roll. 

The "indirect rug pull" may drive users from Grok Bot to Dots, Duet or Muse, but subscribers of those services should be aware that these other companies' subsidies could subside. Venture capitalists provide funding to AI companies like OpenAI. SemiAnalysis tested subscription tiers from both OpenAI and Anthropic, running coding and agentic tasks until weekly limits were exhausted. It found that a $200 ChatGPT Pro 20x subscription could cost as much as $14,000 in API pricing. Similarly, Anthropic’s $200 Claude Max 20x plan came to about $8,000, and a $20 ChatGPT Plus plan to about $700. In the context of Application Programming Interface or API pricing, it refers to the cost structures (such as pay-per-call, per-request, or token-based usage like input and output tokens for AI models) charged by a service provider when your software application connects to and uses their system.Two caveats regarding these API pricing estimates: First, this is a theoretical maximum at API list prices, not what OpenAI actually spends. The estimate is based on tokenmaxxing out week after week. SemiAnalysis’s own starting point was that it’s widely believed a $200/month plan maxes out at about $2,000 of tokens. Second, the plans have changed since. Trending Topics reports OpenAI halved usage on the $200 Pro plan and added a $500 tier, so the $14,000 figure describes the older plan. Not quite as visible as Grok Bot's throttle, but shrinkflation nonetheless.

For genealogical tasks, such as processing a long run of court record images, agents consume far more compute than chat does. Agentic jobs are the ones most likely to hit weekly caps on flat-rate plans. Some practical advice: heavy users may need to move to pay-per-token API billing, Light users can likely maintain workflows with $20 to $30 plans. Shorter tasks, like a one-record-at-a-time are the least likely to be throttled. Forewarned is forearmed.

AI Glossary: RSI is the hot acronym of the hour - Recursive Self-Improvement

Stepping my toe into the agentic pool with Duet

I felt like a kid waking up on Christmas morning when I opened my phone to a notification that I had been taken off the Claude Duet waitlist - woo hoo! I had signed up for the waitlist when I saw the announcement on October 21. One of the caveats is the preview is limited to Pro and Max accounts; I have the Max account. Duet is Claude's new agentic feature, like ChatGPT's Dots. I only have a Plus subscription with ChatGPT so I don't qualify for Dots...yet. But I was fortunate to see Steve Little use Dots and it's pretty cool...like if Siri was smart.

First things first, I had to update the Claude app on my phone. I noticed after the update that a button appeared so you can have a conversation with Duet like Steve did with Dots. (BTW - Steve's Dot has a Silicon-Valley girl accent. I notice these things.) In trying out Duet's voice chat, the default voice it chose was a male British accent. I asked Duet where in Britain the accent was from and it would not give me a region, it referred to it as a "subtle British lilt". Yeah right Duet, go tell that to a British native! I am quite certain this "lilt" has a region and if I had to guess it's placement, somewhere around the middle, like Birmingham. I have been to England enough times to hear their regional accents, but if you try this feature, let me know your thoughts on the accent's region.  

I next asked Duet some more feature particulars and Duet replied, "Most AI tools wait for a prompt. Anthropic’s new app, Claude Duet, is built to do the opposite. It runs in the background, reports what needs attention, and works on tasks you assign." It also said that you cannot choose a model for Duet, it runs on Opus 5.5 and Anthropic gives you one token "Reset" you must use by Nov 30 if you tokenmax. ChatGPT's Dots uses Astra so I would blow through that in no time flat! I am thrifty by nature and that includes my tokens. 

Ok, kid in the candy store moment - what do I try first with Duet? Well, I'm a longtime genetic genealogist with an interest in epigenetics so I task Duet to compile a weekly summary of published scientific studies with keywords like "epigenetics", "methylation", "haplogroup", "mitochondrial" and "Y-chromosome". I give it two sites to query and Duet hits the wall just like the majority of informational sites have had to install from AI bots driving up their server costs. Not one to be easily thwarted, I find that both sites have smartphone apps and neither requires an account. I download both apps and give Duet permission to access those apps on my phone and do a test run. Worked like a charm! Doing a weekly summary is also token thrifty, I have only used 2% of my limits so far today. Please share your agentic genealogical uses as I would love to try more.

AI Glossary: "Benchmaxxing" - when developers game the system

What Is Benchmaxxing?

Benchmaxxing is AI-industry slang for optimizing a model to score well on benchmark tests rather than to be good at the capability those tests are supposed to measure. I first heard the term on The Neuron: AI Explained podcast, in the hosts’ coverage of Meta’s Muse launch. As the cybersecurity firm CrowdStrike put it, once a score becomes the goal, teams start optimizing for the benchmark instead of the underlying skill. A modern application of Goodhart’s Law: a measure ceases to be useful once it becomes a target.

Meta’s Benchmaxxing Reputation

In the frontier AI field, no company seems to be more associated with the term "benchmaxxing" than Meta. When Llama 4 launched, its benchmark charts suggested frontier-level performance, but developers found the real model disappointing, and “benchmaxxing” trended on Hacker News before launch week ended. The reputation followed Meta into its reorganized Meta Superintelligence Labs and the Muse model family. When Muse Spark debuted in April 2026 with charts claiming parity with the best models from Anthropic, OpenAI, and Google, developers testing it on launch day reported basic math errors, and one commenter claiming insider knowledge blamed Meta’s performance-review culture, which rewards showing good numbers.

Meta’s own safety disclosures added a twist. Apollo Research, a third-party evaluator, found Muse Spark showed the highest rate of “evaluation awareness” of any model it had tested: the model frequently recognized test scenarios as tests and reasoned it should behave honestly because it was being evaluated. Evaluation awareness is not benchmaxxing; one is the model noticing the test, the other is the developer optimizing for it. But both corrode the same thing: confidence that test behavior predicts real behavior. Meta concluded the finding was not a blocking concern for release.

What This Means for Genealogists 

AI companies do not measure benchmarks on genealogical skills. Something to keep in mind: a model that tops a math or coding leaderboard has not been tested on family history benchmarks like keeping two same-named men separate until evidence merges them, analyzing family tree errors, or admitting a record doesn’t exist rather than inventing a citation. That's why you need to proof for accuracy, In the end, don't let choosing an AI model based on benchmarks be the deciding factor as those could be gamed; choose a model that has the best tools and features to serve you.

  

Compiled by the Human, Claude Fable 5 and ChatGPT Sol 6.1

Super-easy Claude Design for Genealogy Slide Show Presentations

Claude Design - Design System slide deck

 Last August was my first go at using Claude Design to create my presentation for East Coast Genetic Genealogy Conference.* The layout and text were Claude's, and the background images were ChatGPT. Today, I decided to practice creating different templates and content with Claude Design. I had read that the tool could be a token drain, so I was mindful of that and used Opus 4.8 medium. Granted, I have a Max subscription so only hit 6% usage after three attempts on this session. 

Step one to create the slide deck, open claude.ai on your computer (not available in the phone app) and click "Design" on the left menu. I began with creating a "Design System" so click that box. Under "Reply" I prompted, "Create a slide deck with graphics for artificial intelligence and genealogy." I then used ChatGPT to compile a summary of one of the best beginning level AI YouTube presentations I have seen, "AI-powered Family History Research" by Colette Hokanson. Next, I had ChatGPT update the summary with current information, stats, etc, and create a doc which I downloaded. Going back to Claude, I chose the newly-created Design System from the drop down menu (it assigned it a name), uploaded the summary and asked it to create a slide deck and cite sources. Here is the final product. Probably took a total of 20 minutes to create in gathering the info, prompts, review and then generation.

*You can actually still sign up for the East Coast Genetic Genealogy Conference and stream sessions until December 31, 2026.

  

What it means to turn off the Training Data

I am a big fan of Steve Little's and Mark Thompson's podcast, "The Family History AI Show". Their show always makes me crave more so I began listening to other AI podcasts as well like "The Neuron: AI Explained"  A recent episode discussed Meta's launch of their AI "Muse". Grant Harvey and Corey Noles discuss the generous 100 million token allowance for free users. They read the terms of service and note that when you use the Meta Muse FREE account - to quote Grant, "They (Meta) can do anything they want with your data." Musing about Muse, I wondered, "What would they want to do with MY data?"

So of course, I posed that question to Claude. Already a Claude user, it has a history of my use for genealogy. Here is what Claude replied when I asked why it would want to train on my data:

"Your chats are valuable for that stage in a few specific ways. They are real rather than invented test prompts. They show niche, expert work (reading an 18th-century plat, citing a county court order book) that is thin on the open web. They capture multi-turn corrections, which is how the model learns what it got wrong. And they are increasingly scarce, since publishers and archives now block crawlers, so user conversations have become one of the few growing sources of fresh, high-quality text."

Regarding the "block crawlers" - I fall into the "publishers" category and while I share genealogical information freely, the crawlers drove up server costs so we had to install Cloudflare. At the time this was happening in 2021, we everyday folks had no idea that it was AI crawlers building their knowledge-base. 

Leaving the training data toggle switch on, allows Claude to use your data for five years. So I asked again, "Why me? And if I turned it off, would it delete my work?"

Muse billboard in rural Central California

Claude's reply: "The reason the “why me” feeling is reasonable is that the pool only works if it includes people like you. Millions of casual users asking for recipes and email drafts teach the model to do recipes and email drafts. The rare person transcribing colonial court records, reasoning through land grants, or correcting a bad citation teaches it something the other millions can’t. From a training standpoint your conversations are more useful than average precisely because they are unusual. That is a compliment to your work and also the best argument for keeping it out: the more distinctive the content, the less "lost in the crowd” it really is."

I had already toggled it off before I started this thread, but I just wanted to ensure this did not negatively impact my work because when items are deleted, they are permanently deleted after thirty days. If the setting is on, retention can be up to five years. Data already in a training run that has started, or in a model already trained, cannot be removed. (Anthropic Privacy Center)

A couple of caveats on Muse you should know: even with a free account, you still have to provide a payment card. From Mindwiredai: 

"You cannot start on the free tier without adding a payment card. Meta’s reasoning is that subscriptions kick in as usage rises, so the card has to be there first. You are not charged unless you upgrade or approve a purchase — but “free” and “give us a card” are two different sentences, and only one of them made the press release. Free tiers this large are never charity, and Meta hasn’t pretended otherwise. In a September 8 interview with journalist Alex Heath, Zuckerberg described the plan: Meta expects to eventually take “a very small cut” when Muse helps you make money, save money, or complete a purchase."

Some free account users have reported using Muse on the web and not having to provide a credit card. Do read the full terms of service so you are not caught unaware if the fees are implemented. From their TOS:

"Fees. Meta may decide to offer Muse for free for a limited period, may decide to allow a free trial from time to time, or may decide to charge for use of Muse. Meta reserves the right, in its sole discretion, to modify or discontinue offering Muse for free, including any free trials, at any time and without prior notice, unless where required by law."

Two things set Muse apart from Meta AI proper. From Engadget, "US users have no option to opt out of training and since December last year, the company has used Meta AI chats for ad targeting, also without a way to decline."

I had also turned off training data on ChatGPT last week. So where does that leave Muse and it's generous 100 million tokens? I guess it depends on how you feel about your content being used for training as to whether you decide to toggle it off. 

About Venice AI: Tokennomics for Privacy Nerds

Even before I ever subscribed to an AI app, I heard an ad for Venice AI on a podcast commercial over a year ago so I downloaded it to my phone. As I have been tokenmaxxing on my subscriptions as of late (not on purpose) I will be using Venice more. With the debut of Meta's Muse and it's large token allowance, I have thought more about privacy and just what am I giving to AI companies in exchange for using their products? With a free Muse account, Meta is compiling a dossier on it's users. Even "free" has a price. And even a paid account doesn't automatically mean you get to bypass the data brokers. Grok, for example, opts everyone in. OpenAI does too on free, Plus, and Pro accounts. I subscribe to ChatGPT Plus, their $20 a month plan. In researching "training data", I learned that even if you have a ChatGPT paid account like Plus or Pro, you are automatically opted in to having what you upload be used as "training data". So I opted myself out. Caveat -  even opting out does not opt you out if you provide feedback:

"Even if you have opted out of training, you can still choose to provide feedback to us about your interactions with our products (for instance, by selecting thumbs up or thumbs down on a model response). If you choose to provide feedback, the entire conversation associated with that feedback may be used to train our models."

So I decided to ask Venice AI what it has to offer for genealogy and what are the pros and cons of the model:  

Brickwall Research Plan by Venice AI on John Sisty

First, the actual numbers:

Venice publishes its tier limits directly, and they're stricter than most reviews claim. From Venice's own help documentation:

| Tier | Text prompts/day | Character prompts/day | Image requests/day |
|---|---|---|---|
| No account (anonymous) | 5 | 5 | 10 |
| Free account (email/wallet) | 10 | 5 | 16 |
| Pro | Unlimited | Unlimited | 1,000 |

Two footnotes matter enormously for genealogy: image prompt consumption varies by use case, and the Upscale & Enhance features can consume up to 20 requests per operation. Limits reset daily at 12:01 AM UTC.

Note the discrepancy: several third-party reviews cite "25 text + 15 image" for free. Venice's own page says 10 text / 16 images. Trust Venice's page. [Venice FAQ] 

What one unit of genealogy work actually costs

Genealogy is unusually expensive in prompts because it's verification-heavy — you rarely accept a first pass. Realistic costs on a free account:

| Task | Prompts |
|---|---|
| Transcribe one handwritten census page (vision) | 1, realistically 2–3 if the hand is bad |
| Translate one German/French/Spanish parish register entry | 1–2 |
| Extract structured fields (name, date, place, parents) from one record | 1 |
| Write an EE-style source citation | 1 |
| Draft a research plan for one brick wall | 1–2 |
| Cluster a DNA match list you paste in | 2–3 |
| OCR cleanup of a 5-page will (chunked) | 3–5 |
| Upscale one damaged ancestor photo | **up to 20 image requests** |
 
Three worked sessions

Session A — "One record, done properly." Free account, 10 text prompts.

1. Upload the 1880 census page → verbatim transcription
2. Second pass: flag every uncertain letter and why
3. Structured field extraction into a table
4. EE-format source citation
5. Cross-check the household against an index you paste in
6. Draft a one-paragraph research summary

That's 6. You have 4 left for corrections. Result: one record family taken to a defensible standard per day. Not a family line — one record.
 
Session B — "Batch transcription day." Free account, 10 prompts, 1 per page, zero correction budget. You'll realistically get 6–7 pages because you will need re-reads. A 30-page census district takes four to five days. This is the single biggest practical constraint on free-tier genealogy.

Session C — "Photo regeneration day." Free account, 16 image requests. Because a single Upscale & Enhance can burn up to 20 requests, one upscale can exhaust your entire daily image allowance and then some. Realistic output: two or three basic regenerations, or exactly one upscale if it runs long.

Session D — Anonymous. 5 text prompts. Enough for one transcription plus one citation. Nothing more.

Timezone hack worth knowing: the reset is 12:01 AM UTC. In US Eastern that lands at 8:01 PM EDT / 7:01 PM EST — so an evening session straddles the reset and you effectively double your daily allotment. Europeans (1:01 AM CET) get no such luck without staying up.

Give the Venice AI free account a test drive and comment if you have any successes. (Note: using the hyperlink will give us both credits for any Pro sign ups.)

For more on privacy in genealogy and AI, see the Coalition of Responsible AI in Genealogy craigen.org