In partnership with

🧠 Has OpenAI Solved Hallucinations?

Short answer, no, or – at least, not yet. OpenAI’s latest research spells it out clearly. Large language models still produce confident errors, and the company argues a big reason is how current training and evaluations reward fluent answers over calibrated uncertainty.

The proposed fix is not magic, it is better incentives and better tests. That means scoring systems that punish confident wrong answers more than honest “I am not sure.” 

🔍 What Has Improved

New reasoning models can plan better and cite steps, which helps reduce some mistakes in narrow settings. Retrieval can ground answers in sources. Tool use like browsing can verify details. But when tools are restricted, OpenAI’s own safety notes show some newer reasoning models hallucinate more than earlier ones. Progress is real, perfection is not.

🧪 How OpenAI Says It Will Tackle This

OpenAI is pushing evaluation shifts and training tweaks. Think rewards for admitting uncertainty, stronger grounding, and self checking loops that test one answer against another. External reporting echoes the theme. The goal is to raise the floor of reliability, not to claim a cure. Independent coverage and explainers make the same point, the problem persists even as models get smarter.

⚖️ What It Means for You

Treat confident answers as leads, not verdicts. Ask for citations. Use a second pass with a new prompt that asks the model to challenge its own claims. When the task is high stakes, add retrieval, or check the primary source yourself. If a model tells you it is unsure, that can be a feature, not a bug. The honest “I do not know” is progress too.

Bottom line: OpenAI has not solved hallucinations. It is building systems that fail more gracefully, and that is what matters today. 

Kickstart your holiday campaigns

CTV should be central to any growth marketer’s Q4 strategy. And with Roku Ads Manager, launching high-performing holiday campaigns is simple and effective.

With our intuitive interface, you can set up A/B tests to dial in the most effective messages and offers, then drive direct on-screen purchases via the remote with shoppable Action Ads that integrate with your Shopify store for a seamless checkout experience.

Don’t wait to get started. Streaming on Roku picks up sharply in early October. By launching your campaign now, you can capture early shopping demand and be top of mind as the seasonal spirit kicks in.

Get a $500 ad credit when you spend your first $500 today with code: ROKUADS500. Terms apply.

Holiday performance starts now.

🌐 Google Says the Open Web Is in Rapid Decline

In a recent court filing, Google described the open web as already in rapid decline. The company argued that breaking up parts of its ad tech would make things worse for publishers who rely on open web ads.

This is a notable contrast with public reassurances that search traffic remains stable. The filing and follow up reporting add weight to what many publishers say they are experiencing. 

🔎 What Changed

Google has pushed AI features that answer questions inside search, which can reduce clicks to websites. Recent coverage shows some outlets reporting steep traffic drops, while Google leaders say overall referral trends are steady. The court language matters because it is a formal admission about the open web environment, even if a spokesperson framed it as specific to open web display ads.

➗ What the Data Says Now

Conflicting numbers are fueling the debate. Publisher groups report Google’s AI Overviews correlate with 1–25% declines in referral traffic, suggesting summaries are intercepting clicks that used to go to sites. Pew’s user-level analysis echoes the pattern: when an AI summary appears, people click through far less often (8% of visits) than when it doesn’t (15%). 

Meanwhile, SEO datasets show AI Overviews are spreading across queries—Semrush and BrightEdge both find steady growth in how often summaries appear—while BrightEdge also notes impressions rise but click-through rates fall.

Google’s rebuttal: its “rapid decline” line referred specifically to open-web display advertising, not the entire web, and it says AI still drives billions of clicks. The split screen is real: many publishers see drops; some analytics shops and Google emphasize stability or Discover offsets. 

📈 Why It Matters

If fewer users click through, the economics of independent sites get harder. Expect more licensing deals, paywalls, and experiments with in-page experiences. For readers, this could mean fewer free sources and more summaries at the top of search. For creators, it is a cue to diversify, build direct audiences, and track how AI surfaces compress your work into answers. The web is still here, but its shape is changing.

Bottom line, when the platform that routes attention says the open web is shrinking, believe that signal and plan for it. 

This Week’s Scoop 🍦

🧾 Intel overhauls product leadership as chief of products exits

🧪 AI model predicts flu vaccine strains more accurately than WHO in pilot tests

🎥 OpenAI to debut AI movie “Critters” at Cannes Film Festival

🛡️ Microsoft offers free Copilot services to US government workers

🤖 Sam Altman says bots are making social media feel fake

🚀 Space Force plans to weave AI into everyday operations

🎯 Weekly Challenge: Build an AI Truth Loop for Everyday Life

Until AI hallucinations are fully solved, here’s a quick way to ensure maximum truth seeking with the LLMs you currently use:

📋 Step 1: Save a starter prompt
“Act as a careful checker. I will paste a claim. Return three things. One, a short search plan. Two, the strongest counter possibility. Three, the single primary source I should read.”

🔁 Step 2: Run two passes
Paste the claim, for example a health tip in a group chat or a viral stat. Use ChatGPT, Claude, Gemini, or Perplexity. Ask for a first answer. Then paste the same claim again with a new instruction, “Challenge your first answer, show me where it could be wrong.”

🔍 Step 3: Verify one thing yourself
Open the primary source the AI suggests. Read the title, the date, and one methods paragraph. If it does not match the claim, stop sharing the claim.

💾 Step 4: Save the winner
Copy the best short explanation into your notes with the source title and date. Name the note “Truth loop” and add to it over time. This is your pocket reference when the same claim appears again.

🗓️ Do this a few times this week. You will not catch everything, but you will catch enough to make your feed and your decisions better.

Want to sponsor Jumble?

Click below ⬇️

Will we finally get rid of pesky AI hallucinations, and have you seen signs of the web’s decline in your daily use? We’d love to hear your thoughts. See you next time! 🚀

Stay informed, stay curious, and stay ahead with Jumble!

Zoe from Jumble