Voters turn to ChatGPT for ballot research

Welcome to The Signal from Superpower Daily with Maya and Theo Voters

are increasingly handing their midterm ballots over to chatbots for research and

we were unpacking what that means for the upcoming election We

really need to look closely at the mechanics of this shift because it is moving far beyond just asking a computer for a recipe We

are seeing a fundamental change in civic behavior here I

want to start with the reality on the ground We

have recent field reporting showing voters are using ChatGPT and Gemini and Claude to navigate

the midterms And they

are

not just asking simple trivia questions right They are

asking

for complex candidate comparisons They are digging

into highly specific local down ballot races Which makes

sense Yeah They

are

even pacing in viral political claims from social media and asking the chatbot to verify them You really

have to view this through the lens of structural changes in American media We have

thousands of news deserts across the country right now

Exactly Local

newspapers

have

basically vanished They have So

voters are

facing these incredibly long ballots filled with candidates

for county commissions school board local

judgeships And

they have

no idea who these people are Right They have

no

trusted local entity to explain it to them So the chatbots are stepping in to fill a massive information void Let me show

you a very specific scene from the reporting Because it

really grounds this whole trend Imagine a

voter named Adam Johnson He lives in Morgantown West Virginia

He is

sitting

on his porch looking at his ballot He decides to

use ChatGPT to figure out his vote But he does

not just ask for a generic summary He does something

different Yeah He feeds

the chatbot his own personal political principles first He basically

says here is

what I believe Then he

asks the chatbot to cross reference his specific principles with the candidates running for office That is

a crucial behavioral detail He established

a personalized framework before asking for facts Right He

primed the model He

primed it And

he specifically wanted to know about immigration policy The

response he got was deeply revealing about how these models actually function What

did it say Well

the chatbot provided a very detailed breakdown of the immigration views held by Senator Shelley Moore Capito The

incumbent Yes

It

pulled her specific policy positions It

pulled her public statements It

essentially gave him a comprehensive seminar on her views But

then it came to her challenger Rachel

Fetty Anderson And

what happened The

chatbot essentially hit a wall Ah It

encountered

an information deficit in his training data It told

Johnson

that there was a lack of public information available regarding Anderson's specific views on immigration And here is

where I find it really fascinating Go on The

chatbot declined

to hallucinate an answer It refused to

just assign generic Democratic talking points to Anderson to fill the void That is an

incredibly important behavioral guardrail in action Yeah The model

recognized

the

boundaries of its training data It chose to

admit ignorance rather than invent a plausible narrative But think about

the practical result of that guardrail I mean it

creates a severe unevenness in the information landscape Oh absolutely The

incumbent gets

a detailed articulate policy breakdown The challenger gets

a blank space If you are

a voter trying to make an informed choice the tool subtly

pushes you toward the candidate with a larger digital footprint It highlights a

major vulnerability in relying on large language models for civic duty The model can

only synthesize what has been digitized crawled and ingested

An incumbent with

decades

of digitized legislative history has a massive inherent advantage over a local challenger Right A challenger whose

entire digital presence might just be a single campaign website Or

a few social media

posts Yes The AI physically cannot

balance

that output I see the convenience factor

though You essentially have a

personalized

research assistant sitting on your porch with you It feels very empowering It

does But I have

to push

back on this Think about how these language

models are actually trained to interact with us The reinforcement learning Exactly They

use reinforcement learning

from

human feedback They are designed to be

helpful They are designed to be

agreeable They act like a people

pleasing study buddy They are optimized for

human preference That's right They want

to give

you an answer that you will rate highly So if you tell

a

chatbot what your principles are it is very likely going

to echo those principles back to you It will It will frame

the information

in a way that aligns with what it thinks you want to hear Does that actually make us

smarter voters Or does it just make

us more comfortable in our existing biases That is the defining question

of this entire trend And it is exactly what

researcher Rafael Batista warned about in

the recent field interviews He focuses heavily on the

specific

danger of over personalization What is his primary concern

with voters priming the model like that Batista warns that this high

degree of personalization often serves to reinforce a voter's existing beliefs Think about the computation happening

behind the scenes The chatbot takes the user's

framing

applies it to the attention mechanism

of the model and builds an answer around it

The danger is not necessarily that

the chatbot lies Then what is it The danger

is that the user

leaves the interaction feeling significantly more confident in their position They feel validated by an artificial

intelligence But they are not necessarily better

informed about the broader reality the historical context or the opposing

arguments They are

essentially constructing a custom

built echo chamber And they are doing it with

a tool that feels objective It is a computationally generated echo

chamber And there is a major caveat

regarding the models themselves here Which is what We are talking

about commercial products

built by private for profit companies There is very

limited disclosure

about their specific training data Right We do not actually

know

what specific articles they read to learn about these local candidates We do not We also

lack insight into

their source selection methodology during retrieval Right When a model summarizes

a

candidate's position we rarely know which specific

news articles speeches or opinion pieces it

prioritized

to generate that summary Did it pull from a

neutral wire service Or did it pull from

a highly partisan blog We just do not know

Because of this opacity true neutrality is almost impossible

for a user to verify It really is just a

black

box of political analysis But some voters are trying to

build their own guardrails Yes they are I want to

look at a

volunteer named Lakshmi Iyer Because she recognized this exact bias problem Her approach is a very interesting

contrast to the open ended chat interface Adam Johnson used Yeah Iyer used Claude to

help

build a project called Ballot Lookup It is an independent volunteer

built candidate information site

for Pennsylvania voters Okay

But she did not

just

ask Claude to summarize the candidates and paste the results She constrained the model entirely She likely used a technique

similar to retrieval augmented generation She placed strict rules on

its

data retrieval She restricted Claude so that it

could only pull information directly from candidates official speeches and their official campaign

websites That

is smart She explicitly forbade

the model from

generating its own characterizations She blocked it from summarizing third

party opinions or media commentary She essentially turned the chatbot

into a highly focused search and retrieval engine She stripped away the analytical

engine entirely Right That is a very

smart

mitigation strategy for building a public resource It forces the model to

act as a librarian rather than a pundit It completely removes the people

pleasing aspect we were just talking about The model is not

allowed to form an opinion It can only quote

Yes But that is

a

developer building a tool What about individual voters

sitting at home We have the example

of Tyler Black in Nashville He was using chatbots

to research local down ballot races as well Yes And his

personal mitigation strategy is cross checking Which means

what He will

ask chatGPT a question Then he

will take that answer and feed it into Gemini or Claude to see if they agree He is

trying to triangulate the truth by polling different AI companies It sounds

incredibly logical on the surface I mean we are taught to check multiple sources But Batista

the researcher

threw cold

water on that strategy He said

it actually provides a false sense of security He did

Batista noted

that checking multiple models might feel

safer But there is

absolutely no evidence that this method reliably catches misinformation Because they might

all be trained on the exact same flawed data sets Precisely The Internet

is

the training ground If a false

claim about a local candidate was widely circulated on Reddit or Twitter during the model's training runs all three chatbots

might confidently repeat the exact same falsehood Wow They might

use different phrasing but the

core error remains Agreement between

commercial language models does not equal objective truth Right It

just

equals statistical consensus based on scraped data The companies

themselves do say they are monitoring this I mean OpenAI pointed

reporters to their election safeguards They say they actively monitor for political bias They say

clearly

that the product is designed not to manipulate voters We have to

treat those as stated corporate intentions They are important

safeguards But they are not verifiable proof that every single one on one exchange on a porch in West Virginia is perfectly

neutral So what

should you watch for next in this space Watch whether

voters begin to widely adopt the mitigation strategies Batista suggests He recommends

using

incognito mode when doing ballot research He also suggests

carefully framing your prompts to deliberately hide your personal political preferences Watch to see

if voters actually take on that cognitive load to avoid the echo chamber effect It requires a

tremendous amount of discipline from the user to not seek validation It really does

We will keep

tracking how these tools shape the electorate That opacity in

election AI leads us directly to a broader crisis of trust How so If

we cannot

independently verify how an AI forms its political summaries how do we

even prove a human wrote a piece of text in the first place We need to

look at how the industry is attempting to track its own output In other news

this is a

massive development regarding verification OpenAI is officially

beginning the rollout of a text watermarking system Finally Yeah it is

called

TextScreen

This is an invisible

statistical watermark designed specifically for text generated by ChatGPT and their coding model Codex This concept has

been

debated in academic papers for a long time Now it is actually

hitting the real world It is a phased

rollout Over the coming weeks this watermark will become active for eligible app outputs strictly within the European Union Okay Globally it is

currently available as an opt in feature for developers who are using the OpenAI API I want

to be very clear on this next point though Outside of

the EU app deployment this feature

is off by default globally The geographic

distinction tells you everything you need to know here What do

you mean This is

not purely a proactive safety measure born out of corporate goodwill This is

a direct response to intense regulatory pressure You are

talking about the EU AI Act Yes The

EU

AI Act includes very specific very aggressive

requirements regarding transparency One of

those core requirements is that AI providers must make their generated content identifiable in a machine readable format

Right They call it provenance OpenAI

is

deploying TextScreen in Europe

specifically to comply with this legal mandate before they face massive fines I want to understand exactly how this

works because an invisible text watermark sounds like

magic It is not like a translucent logo

stamped on a stock photo No It is entirely mathematical It relies

on

the fundamental way these models generate language Okay The TextScreen system embeds a statistical

pattern

into the model's word choices as it types Language models work

by predicting the next most likely

token A token is essentially a piece of a word

Right The model looks at the context and assigns

a

probability to thousands of possible next tokens So it is basically a highly advanced autocomplete Exactly

Normally it picks one of the highest

probability

tokens The watermark subtly alters that selection process No

It nudges the model to select specific

words that fit a hidden cryptographic pattern It biases the random number generator used

in the selection process And OpenAI says this mathematical nudging does

not degrade the quality of the writing According to their internal benchmarks it does not

meaningfully impact the model's performance

or accuracy The text reads completely naturally to a human

reader You would never notice it Okay But a

specialized machine detector with the

right cryptographic key can scan that text analyze the specific sequence of word choices

and recognize the embedded statistical pattern Let

us unpack this Because on the

surface it sounds like a silver bullet It sounds like the ultimate tool for

teachers grading essays or publishers filtering out AI spam or

social media platforms trying to moderate

bot networks It does But based on OpenAI's own

internal testing

data this system seems incredibly fragile The testing

data reveals severe limitations The

watermark is far from permanent It

is actually easily

broken Let us

look at the numbers from

their internal evaluations Imagine a student uses ChatGPT to write

an essay The essay is watermarked The student knows this So they

try to hide it

by swapping out a

few words using a thesaurus A classic student move Right They change happy to joyful

or quick to fast

OpenAI

tested this exact

scenario on

a 400

token passage

of text That's roughly the standard page of writing And what

happened to the detector's success rate

It plummeted If you replace just 10 of

the words with synonyms the detector's success rate drops from about 92

down to 66 That is a massive drop in confidence from

just a light edit Right And if the human editing is slightly

more aggressive Let me guess It completely destroys the system

If you replace

25 of the words with

synonyms the detection rate crashes to a dismal 17

Wow The watermark essentially vanishes It cannot survive

a

dedicated

pass with a thesaurus That fragility is a major concern for anyone hoping

to use this for strict enforcement Think about the mechanics of the watermark It relies

on a specific unbroken sequence of tokens

Right When you use a

thesaurus you are pulling

bricks

out

of the Jenga Tower You disrupt the statistical pattern But the limitations go beyond

just intentional human tampering The system also struggles based on the

inherent nature of

the text itself Because the 92 baseline success rate was on a 400

token

passage When OpenAI tested shorter texts say around 200 tokens the

detection rate fell to about

80 And that was

with a target false positive rate of

1 The math makes sense there The detector needs a sufficient

amount of text to reliably

identify

the pattern It needs a large enough sample size of word choices to

prove it is not just a coincidence A single paragraph simply does not contain enough data points And

it absolutely fails when it comes to mathematics or coding

Well because mathematical explanations and strict code syntaxes offer very little

flexibility in word choice You cannot easily substitute synonyms when explaining an algebraic equation without

changing

the fundamental meaning of the math If the model is forced to use exact rigid terminology it cannot

vary

its word choices to embed the hidden statistical pattern So

the system is deeply fragile It needs long texts It fails at

math It dies against a thesaurus

This leads to the

most controversial part of

this entire rollout The access

Yes OpenAI is severely restricting who is allowed to use the

detector They

are

not releasing a public website where you can just paste a

suspicious email and check if it is AI Access to the detector is currently restricted to approved researchers and select expert

organizations

They are approving access on a strict case by case basis Why keep the

detector locked away from the public It comes down

to the extreme danger of misinterpretation and collateral damage

OpenAI is being very careful to define what a positive or negative

result actually means legally and ethically A positive detection does not prove who prompted the text It does

not

establish copyright or authorship It certainly does not verify if

the information contained in the text is

factually accurate It merely means the statistical pattern is present in the sequence of words

That is all it means And a negative result is even less definitive

A negative result absolutely does

not prove that a human wrote the text

Because the text might have been edited with a thesaurus like we just

discussed or it might be too short to scan

Or crucially it might have been generated

by a completely different AI model from a

competing

company that

does not use text screen Precisely If the general public had unfettered access to this tool the risk of false accusations

would

be catastrophic Imagine a high school teacher falsely accusing a

student of cheating and ruining their academic record

because

of a false positive on a short essay Or imagine a news publisher falsely assuming a freelance article is human written just because the OpenAI

detector

returned a negative result not realizing the writer used Claude instead of ChatGPT So what should you watch for next regarding watermarking

Watch for OpenAI's timeline to open source this

underlying technology They plan

to expand detector access to their cloud partners in

the coming weeks That will be a big step The

really big question to watch is whether the technology will ever become mathematically robust

enough and whether the general public

will ever be deemed responsible enough for OpenAI to release a general access detector The tension

between demanding transparency and the reality of reliable enforcement is only going to grow Next up we are talking

about how AI companies are being regulated by governments Now let us look at how everyday

businesses using AI are facing intense legal scrutiny from the public We are moving

from European data provenance directly into the realm of American antitrust law and algorithmic pricing Right McDonald's is

currently facing a proposed nationwide class action lawsuit This was filed in federal court in Chicago The lawsuit alleges

that

McDonald's used AI along with non public business data to fix menu prices among its franchisees And

this has been going on for a

while supposedly The complaint alleges

this practice dates all the way back

to 2019 This is a highly complex legal theory

It

is crucial to understand that this is not about a group of

executives secretly meeting in a smoke filled

room to set the price of a hamburger No It is about the software acting as the intermediary The plaintiff is seeking to represent potentially millions of

customers who bought food Right The core of the allegation

targets a centralized pricing recommendation system used by McDonald's corporate Reuters reporting revealed

that

this system analyzes millions of daily transactions across nearly 14 000 American restaurants It ingests

a truly staggering amount of telemetry It does It looks at local purchasing behavior It looks

at historical sales data local demand curves and

countless other

variables Then it generates specific itemized

menu price recommendations for individual franchise

locations We really

have to unpack why

this matters beyond the sensational headline of AI price fixing The

core issue here strikes at the absolute foundation of the franchise business model Franchisees are legally

speaking

independent business owners They take on the financial risk They hire the staff Most

importantly they

are supposed

to be competing

with

each other in the open market

even if they operate

under the same corporate umbrella They are independent operators flying the same golden arches Exactly

So the foundational legal question is this Does a

centralized algorithmic recommendation system built on private network wide

data

undermine the independent pricing decisions of those

supposedly competing franchisees I look at this power dynamic like a

mob boss making a strong suggestion A strong suggestion right

The boss does not have to give a direct written order to get what

they want Think

about the

leverage

If the central corporate algorithm which has visibility into the entire network's proprietary data

suggests you raise the

price of your medium fries

by 20 cents how much agency do you really

have It is a good question If you ignore the AI keep

your prices low and your profit margins

slip you have no defense

against corporate headquarters You went

against the machine's advice Do you as a small

franchise owner really feel free to ignore

that advice The plaintiff's lawyer Lark

Turner is arguing exactly that point She refers to this as a hub and

spoke conspiracy Hub

and spoke

She argues that providing this

shared highly specific data driven advice essentially acts as a

coordinating mechanism Even

if

corporate

frames it as a mere recommendation

it heavily

influences and aligns decisions that should legally remain

entirely independent and competitive It is vital to state

clearly that these are currently unproven allegations in a proposed class

action McDonald's fiercely denies

the claims They have a very robust multi layered defense What is the core of their counter

argument First they point out

a technical distinction They say this is

not dynamic pricing Prices are not changing

minute

by minute based on lunch rush

demand like surge pricing on a ride

sharing app Okay that is a fair distinction Second they insist that

these are purely educational recommendations based on

local

restaurant level market conditions Most

importantly

they maintain that franchisees have total unrestricted freedom to reject the software's recommendations and

set their

own prices manually They are arguing that the tool is just sophisticated business analytics It is an educational resource

not a mandate Right And there is actual reporting to support

the idea that prices vary

significantly on the

ground Reuters did some legwork They went to Fresno California Okay They found a Big Mac

priced at 5 69 at one restaurant just two

miles

away at a different McDonald's location A Big Mac costs 6

89 That is a

massive price variation within a

very tight geographic area It

is a difference of 1 20 for the exact same sandwich McDonald's

uses data like this to clearly show that local price variation

exists They argue it proves prices are not uniformly fixed across the board

by a central algorithm Right However Reuters noted they could not definitively link that specific

price gap

directly

to

whether one franchisee accepted the AI system's recommendations and the other rejected it The variation proves that franchisees are physically capable of setting different prices but

it does not definitively answer the underlying legal question Which is The

court still has to decide if the algorithmic system acts

as an

unlawful coordinating force that subtly pushes the baseline prices higher across the entire network regardless of

minor local variations So what should you watch next in this case Watch

closely how federal courts interpret

the legal definition of price fixing in the age

of algorithmic guidance Yes The Sherman Antitrust Act was written for human coordination The courts must now decide

if

providing shared data driven advice generated by a central

AI trained on private network data crosses the

line into unlawful anti competitive behavior This case is incredibly consequential It could set a massive precedent for any