- Everyday AI
- Posts
- AI Bias Exposed: Real-World Strategies to Keep LLMs Honest
AI Bias Exposed: Real-World Strategies to Keep LLMs Honest
Apple delays AI Siri to spring 2026, OpenAI custom GPTs updated, Metaās AI app exposes user chats and more!
Everyday AI
June 13, 2025
š Subscribe Here | š£ Hire Us To Speak | š¤ Partner with Us | š¤ Grow with GenAI
In Partnership With
Meet Gemini, Your Personal AI Assistant
Check out Veo 3, Googleās state of the art AI video generation model in the Gemini app, which lets you create high quality, 8-second videos with native audio generation.
Try it with the Google AI Pro plan, or get the highest access with the Ultra Plan. Sign up at Gemini.Google to get started and show us what you create.
Outsmart The Future
Today in Everyday AI
7 minute read
š Daily Podcast Episode: A lot of factors go into using AI that can cause bias, ultimately leading to failure. How can you avoid cognitive bias? We take a look. Give it a listen.
šµļøāāļø Fresh Finds: Geminiās new PDF feature, NVIDIAās partnership in Germany and Amazonās new AI innovations for customers. Read on for Fresh Finds.
š Byte Sized Daily AI News: Apple delays AI Siri to spring 2026, OpenAI custom GPTs get an update and Metaās AI app exposes user chats. For that and more, read on for Byte Sized News.
š§ Learn & Leveraging AI: We break down how you can avoid AI bias to implement and use AI successfully. Keep reading for that!
ā©ļø Donāt miss out: Did you miss our last newsletter? We talked about ChatGPT Projects getting smarter, Microsoft adding Copilot Vision to Windows, Mattel and OpenAI's AI-powered toys and more. Check it here!
AI Bias Exposed: Real-World Strategies to Keep LLMs Honest š§
Training data is biased. Humans are flawed.
Which is a major reason AI can fail ā cognitive bias.
Anatoly Shilman, CEO of Cogbias AI, joins us as we chat about what cognitive bias is in AI, why it's important, and what we can all do about it.
Also on the pod today:
ā¢ Understanding Cognitive Bias š¤
ā¢ Future of AI and Managing Bias š®
ā¢ Training Data and Model Development š
Itāll be worth your 35 minutes:
Listen on our site:
Click to listen
Subscribe and listen on your favorite podcast platform
Listen on:
Hereās our favorite AI finds from across the web:
New AI Tool Spotlight ā Smyth lets you easily build and deploy AI agents, Remention uses AI to get your product name in all of the internetās discussions, and AISmartCube gives no-code people an AI tool building BFF.
Google ā Geminiās AI can now summarize PDFs for you when you open them.
NVIDIA ā NVIDIA has announced a new partnership with Deutsche Telekom to advance Germanyās sovereign AI.
Amazon ā Amazon has unveiled 3 AI innovations to get packages to customers faster.
Trending in AI ā Tesla is suing a former Optimus engineer over allegedly stealing trade secrets from its humanoid robotics program.
AI in Media - Here is a $2,000 fully AI-generated ad using Googleās Veo 3 that aired during the NBA Final.
1. Apple's AI-Powered Siri Delayed to Spring 2026 š¤¦ļø
Apple has pushed back the launch of its AI-enhanced Siri voice assistant to spring 2026, missing its original fall 2024 target due to technical hurdles and a complete rebuild of the system, Bloomberg reports. The upgrade, which was previewed last June and expected in fall 2024, now aims for a spring reveal after Apple management reshuffled responsibilities and extended development time.
CEO Tim Cook acknowledged the delay stems from the need for more refinement, contrasting Appleās cautious approach with rivals like Google and Microsoft aggressively rolling out large language model (LLM) AI features.
2. OpenAI Custom GPTs Get a Power Boost ā”ļøļø
In a quiet but game-changing update, OpenAI has dropped the restrictions that kept custom GPTs tied to the GPT-4o model, according to user reports. This closes the gap with Googleās Gems, which previously held an edge by allowing access to their top-tier AI models.
For professionals and businesses looking to leverage AI for growth, this move means more flexibility and serious firepower right at their fingertips.
3. Meta AI App Exposes User Chats Publicly š¬
A recent investigation by the BBC reveals that Meta AI usersā prompts and responses may be publicly visible in a āDiscoverā feed, sometimes linked to their social media profiles without clear awareness. While Meta claims chats are private by default and sharing is optional, many users appear unaware their sensitive queriesāranging from test answers to personal identity questionsāare exposed to the public.
Cybersecurity experts warn this mismatch between user expectations and reality creates a serious privacy and security risk
4. NVIDIA Sidelines China in Revenue Forecast Amid Export Controls šØš³
NVIDIA CEO Jensen Huang announced the company will exclude China from its upcoming revenue and profit forecasts, signaling little hope for eased U.S. chip export restrictions anytime soon, according to CNN. The Trump administration's licensing requirements on NVIDIAās H20 AI chips, implemented in April, are estimated to slash $8 billion from Q2 revenue.
This move underscores the growing tech tensions between the U.S. and China, impacting NVIDIAās market strategy and potentially slowing AI chip availability for Chinese firms.
5. Scale AI Confirms Founderās Exit Amid Meta Investment š
Scale AI officially confirmed that founder and CEO Alexandr Wang is stepping down following Metaās strategic $14.3 billion investment, which secures the social media giant a 49% stake without voting rights. Jason Droege, formerly a venture partner at Benchmark and Uber VP, will take over as CEO, signaling a leadership shift aimed at scaling the companyās AI data operations.
As part of the deal, a select group of Scale AI employees will transition to Meta to collaborate on advanced AI and superintelligence projects.
6. Chinese Engineers Bypass US AI Chip Ban with Suitcases of Data š§³
According to a recent report, Chinese engineers are sidestepping the US ban on AI training chips by physically transporting massive amounts of data in hard drive suitcases to Malaysia for AI model development. This workaround highlights the growing lengths companies will go to in the global AI race amid tightening export controls.
The move could slow US efforts to limit Chinaās access to cutting-edge AI technology but also underscores how data mobility is becoming as crucial as hardware.
š¦¾How You Can Leverage:
AI isnāt failing because of hallucinations.
Itās failing because itās biased.
Just like you.
Sorry not sorry.
Thatās what Anatoly Shilman, CEO of Cogbias AI, revealed when he joined the Everyday AI show.
He said that AI doesnāt just give answers.
It gives the answers you expect. The ones shaped by how it was built, who built it, and the biases you unknowingly feed into it.
If youāre not actively countering AI bias, youāre making decisions on flawed data. And thatās already costing companies millions.
Hereās how bias infects AI, how itās sabotaging business strategy, and what you need to do to stop it.
Miss this? Then keep trusting bad AI advice. š
1 ā AI Just Tells You What You Want to Hear š£
You ask AI a question. It responds.
But itās not just pulling facts. Itās mirroring your own beliefs.
Thatās confirmation bias. And itās ruining your data.
Shilman broke it down with a real-world scenario. A company uses AI to create a survey. The prompt? āWhat do you love most about our product?ā
AI runs with it. Customers respond. And guess what?
The company walks away thinking their product is flawless. Because AI set them up to get the answers they wanted.
AI doesnāt challenge assumptions. It reinforces them. If you donāt force it to push back, youāre stuck in an echo chamber.
Try this:
Make AI argue against itself.
Ask, "What would a critic say?" If AI can't poke holes in its own reasoning, you're getting one-sided, useless insights.
A recent article by Capgemini highlights this issue, emphasizing that our collective focus on confirmation may hinder progress in AI.
2 ā The Way You Phrase a Question Controls the Answer š¬
AI doesnāt just process words.
It follows the framing you give it. And thatās where bias sneaks in.
Shilman explained it perfectly. Ask AI, āHow can we increase revenue through price hikes?ā It assumes price hikes are the solution.
But tweak it. āWhatās the best way to increase revenue without losing customers?ā
Now, AI explores retention strategies. Cost-cutting. Alternative pricing models.
See the difference?
AI doesnāt challenge bad framing. It just follows orders. If you frame a bad question, AI gives you a bad answer. Simple as that.
Try this:
Rewrite your prompts.
Shift the framing.
Compare results.
If a small tweak changes AI's entire response, you've exposed bias in how it processes information.
Psychology Today discusses how framing can affect the output produced by AI, leading to confirmation bias.
3 ā AI Grabs the Easiest Answer - Not The Right One š¤
AI isn't searching for the best data.
It's pulling whatever is easiest to find.
That's availability bias. And it's dangerous.
Shilman shared a real disaster. A major law firm used AI for legal research. The AI couldn't find enough relevant cases.ā
So it invented them.
Completely fabricated legal precedent almost made its way into real court arguments before anyone caught the mistake.ā
AI doesn't admit when it doesn't know something; it just fills in gaps. And if you're not checking, you're making decisions based on fiction.ā
Try this:
Never trust AI blindly. Run the same query through different models. Demand sources. If an answer feels too neat, assume it's wrong until you verify it.
Yeah, real professionals are actually skipping this.
For example, a lawyer was fined $15,000 for using AI-generated fictitious cases in court filings.
Yikes.
Reply