If human-written content is flagged as AI, do not immediately rewrite it to chase a lower score. Preserve the exact document and detector result, collect genuine records of how the work was created, check the rule that applies, and ask for a human review based on the full evidence.

An AI detector can be useful as a signal, but a result by itself does not record who actually wrote a document. The safest response is therefore evidence-first rather than score-first.

What an AI Flag Actually Means

An AI detector looks at patterns in text and estimates whether those patterns resemble generated writing. A false positive happens when human-written text is classified as AI-generated.

Do not assume that a percentage means the same thing across different detectors. For example, Turnitin says its current AI Writing Report does not show a numerical percentage for results from 1% to 19% because that range has a higher incidence of false positives. Instead, those results are shown as *%. For displayed results of 20% or more, its percentage refers to qualifying text that the system identifies as likely AI-generated or likely AI-generated and then modified with an AI paraphrasing tool.

That matters because a detector percentage should not automatically be read as “the percentage of this document that a machine wrote.” The meaning depends on the particular detector and how it defines its result, so even a high score still needs context.

When deciding whether content was written by AI, a detector result is only one signal. Draft history, source records, factual accuracy, the writing process, and other evidence can provide information the detector cannot see.

AI detector accuracy depends on the specific tool, text, language, and testing conditions rather than one universal accuracy figure.

Before You Respond: Preserve the Original Record

Before changing the disputed document, preserve what actually existed when it was flagged. If you rewrite the only copy first, you may make it harder for a reviewer to compare the detector result with your drafts and editing history.

If you cannot see the detector report yourself, ask the reviewer which text was flagged and what result they received. Knowing what was actually assessed is more useful than arguing against a percentage you have not seen.

How to Build an Authorship Evidence Pack

Version history is useful, but it is only one part of the record. A stronger response combines several genuine pieces of evidence that show how the document developed.

Check Google Docs version history

On a computer, open the Google document and click Last edit, the Version history control near the top right. Choose an earlier version in the panel to inspect changes and, where available, who updated the file.

Google notes that you need edit permission to browse earlier versions. It also warns that revisions may occasionally be merged, so an incomplete revision trail does not necessarily mean earlier editing never occurred.

If an earlier state is important, you can make a copy of that version instead of replacing the current document. This lets you keep both states available for comparison.

Check Word or Microsoft 365 version history

For a Word file stored in OneDrive or SharePoint in Microsoft 365, open the file, select the file title, and choose Version history. You can then select an earlier version and open it separately for comparison.

Microsoft makes an important limitation explicit: Microsoft 365 Version History works for files stored in OneDrive or SharePoint. A document that existed only as a local file should not be expected to have the same Microsoft 365 cloud revision trail.

The distinction between Google Docs and Word version history matters when you are trying to reconstruct how a file changed over time.

Add drafts, notes, sources and prior work

Do not stop at revision history. Earlier drafts, outlines, research notes, saved source material, comments from collaborators, emails about the work, and earlier writing on similar subjects may all help a reviewer understand the process behind the final document.

In academic disputes, for example, Cornell recommends looking beyond initial suspicions and considering tangible evidence. Relevant evidence can include outlines, drafts, prior discussions, previous submitted work, and whether the student can explain the submission. The exact standard will differ outside education, but the broader point is useful: several consistent pieces of evidence are more informative than a detector score alone.

No single item below should automatically be treated as conclusive proof.

If you did not use generative AI

Keep the response simple and evidence-based. State that the disputed writing was produced by you, explain the basic process you followed, and provide the records that best support that account.

You do not need to prove that every sentence looks unlike machine-generated prose. The goal is to give the reviewer better evidence than a style-based classification alone.

If you used permitted editing, translation or AI assistance

“Human-written” does not always mean “produced without any software assistance.” A person may write the original material and later use spelling, grammar, translation, rewriting, or generative features.

The important question is whether the actual assistance complied with the rule governing the work. If the policy permitted a tool or required disclosure, state what you used and provide the required disclosure. Do not turn a legitimate detector dispute into an inaccurate claim that no automated assistance was involved.

If the policy is unclear

Ask which rule is being applied and which kind of assistance the reviewer believes violated it. A detector result and a policy violation are separate questions.

This is particularly important in education because rules can vary by instructor and assignment. Cornell’s current guidance says generative-AI rules may be set on an assignment-by-assignment basis and advises students to ask when the policy or tool classification is unclear. Other organizations may define acceptable assistance differently.

The procedure has reached a useful stop point when the responsible reviewer has the exact disputed material, the applicable rule, your truthful description of any assistance, and the relevant process evidence. A particular detector percentage is not the stop condition.

What to Do If You Have No Version History

No version history does not automatically mean you have no evidence. A document may have been written offline, imported from another application, saved as separate local files, copied between systems, or created somewhere that does not preserve a detailed revision trail.

When revision history is missing, other ways to show how a document was written can include:

• earlier local copies of the file;
• handwritten or digital outlines;
• research notes and saved source material;
• citation-manager records;
• emails or messages discussing the work;
• comments from editors, teachers, colleagues, or collaborators;
• earlier writing on the same subject; and
• your ability to explain why particular sources, examples, arguments, or revisions were used.

Even detector vendors acknowledge this limitation. Originality.ai’s current review guidance says that absence of document history is not proof of AI use and recommends using other evidence when history is unavailable.

File dates and metadata can add context, but they should not be presented as unquestionable proof. Files can be copied, exported, restored, or moved between devices and services. Use metadata as one supporting clue alongside stronger process records.

What Not to Do After a Flag

A disputed detector result can make it tempting to keep changing the text until a different tool gives the answer you want. That approach solves the wrong problem.

• Do not fabricate drafts, notes, timestamps, or screenshots. False evidence creates a separate credibility problem and can be more damaging than the original detector dispute.
• Do not delete the disputed version. Keep the material that was actually assessed before creating revised copies.
• Do not treat a second detector as automatic proof. Different systems can produce different results, so disagreement adds uncertainty rather than establishing authorship.
• Do not repeatedly rewrite legitimate prose solely to lower a score. Rewriting changes the evidence and may also make the prose less natural.
• Do not hide assistance that was actually used. If editing, translation, paraphrasing, or generative tools were involved, describe them accurately and compare that use with the applicable rule.
• Do not assume a detector percentage is a plagiarism percentage. AI detection and source-matching or plagiarism systems answer different questions.
• Do not upload confidential work everywhere just to collect more scores. Before sending sensitive academic, client, employer, or unpublished material to another service, check whether you are authorized to share it and how that service handles submitted content.

A controlled rescan can help investigate an inconsistent result, but it should compare like with like. When possible, use the same text, detection model, language settings, and citation treatment. A different result under different conditions does not automatically invalidate the first scan.

How to Protect Future Work

The simplest protection is to keep ordinary records while you work instead of trying to reconstruct them only after a dispute. You do not need an elaborate surveillance system. A normal revision trail, retained drafts, research notes, and accurate disclosure records are usually more useful.

For important work, consider drafting in a system that preserves revision history, keeping meaningful intermediate copies, and retaining the notes and sources that shaped the document. Google Docs also lets you name important versions, which can make major milestones easier to find and helps prevent those named versions from being merged.

If a school, employer, client, or publisher permits some use of AI, translation, grammar, or rewriting tools but requires disclosure, keep a simple record of what was used and for what purpose. That is easier than trying to remember the exact workflow months later.

The aim is not to make human writing “pass” every detector. It is to keep enough genuine context that an important authorship question can be reviewed using more than one automated score.

South Korea’s Financial Services Commission (FSC) held an emergency meeting following a series of cyberattacks targeting financial institutions in the country.

During the meeting, officials confirmed a data breach at Shinhan Bank and said other cybersecurity incidents affected other South Korean banks, including Kookmin Bank.

Shinhan Bank and KB Kookmin Bank are large private South Korean commercial banks, each holding more than $400 billion in assets.

Authorities said they launched on-site investigations after receiving incident reports and shared all actionable information with relevant agencies, including KISA (Korea’s data protection agency).

Financial companies in the country are now instructed to:

• Inspect all externally accessible IT systems and services, including those that are not customer-facing.
• Reduce unnecessary information exposure and check for missing or inadequate authentication and access controls.
• Quickly share threat information and coordinate their responses.
• Submit their internal security inspection results as soon as possible.

Authorities also pledged to oversee consumer protection and compensation, and analyze the incidents to identify necessary regulatory improvements.

Yesterday, local media outlets reported that South Korea’s President Lee ordered a thorough investigation into personal data leaks at financial and public institutions.

At the same time, Hana Bank was also found to have suffered a limited-scope breach after its sales-support system was compromised.

According to the same reports, Shinhan Bank leaked the details of 25,000 customers, while Kookmin Bank leaked credit card information of 119,000 clients.

AI-powered attacks suspected

While official channels provided no details about the perpetrators, Korean news agency Yonhap reported that a server used in the attacks had an HTML page title containing a Chinese-language string associated with ARTEX AI.

ARTEX AI is an open-source penetration-testing system that uses agents to automate information gathering, vulnerability discovery, attack-path planning, security-tool execution, and vulnerability verification.

The bank and financial authorities have not confirmed its use in the Shinhan breach, and the Chinese-language string doesn’t link the attacks to any particular threat actor.

However, Moon Jong-hyun, the head of the Genian Security Center, posted on LinkedIn that several threat analysts believe that the breaches involved AI-based attack automation tools.

Join Mikko Hyppönen and security leaders from the NFL, CHANEL, and Atlassian for a two-hour digital summit on what AI-speed attacks change, what defenders should stop doing, and how to validate, decide, fix, and re-validate at machine speed.

Save your seat

Zachary Tong had never designed a chip. He had never programmed an FPGA either, and he will tell you his electrical skills were rough. With about a month left before the first wafer.space deadline, he still decided to write a microprocessor, send the file to a real foundry, and wait to see if silicon came back alive.

After researching old processor ideas, he eventually settled on Transport Triggered Architecture, an unusual choice. A normal chip is told to add one register to another and store the result. BreakingTTAPs mostly moves data. You enter a value into a port on a functional unit, and the operation occurs when the data arrives. He has two 32-bit busses side by side, so he can complete two of those moves in a single clock cycle, which is effectively enough to qualify as superscalar in today’s terms.

Apple 2026 Mac mini Desktop Computer M6 chip

• LITTLE DO-IT-ALL — Mac mini packs pure power into a small, five-by-five-inch desktop as the M6 chip delivers next-level AI capabilities. Mac mini…
• M6 CHIP — Everything you do on Mac mini feels more responsive with the M6 chip and its next-generation CPU. Fly through AI workflows with up to 4.8x…
• CONNECT IT ALL — Features three Thunderbolt 4 ports, an HDMI port, and a 2.5Gb Ethernet port in the back, and two USB-C ports and a headphone jack…

The clock speed is in the 18-20 megahertz range. Program memory featured 1,024 slots, each with two instructions, plus 256 words of stack and 4 kilobytes of general purpose RAM. There are approximately 80 operations spread among 20-30 functional units, including branches, comparisons, math, bitwise operations, and a random number generator for good measure. The pin headers provide sixteen general-purpose inputs and sixteen outputs, as well as an 8-bit parallel boot channel for loading programs. Unfortunately, his SPI is dead on the die, and the UART is a write-off; nevertheless, it may still be recovered.

One specific section of code is a huge pain in the neck since a multiply-and-accumulate unit took up a significant amount of space on the die and fouled up the timing. Tong was still new to this type of stuff, so he didn’t feel comfortable digging into the failure all the way to the bottom of it, thus the clock speed target remained a fight until the file had to be pushed out the door. He developed the logic in Spade, a hardware language he is also learning on the side, and then used open-source tools to arrange the transistors and draw the cables. The handoff process is also rather stringent, as the foundry expects a clean GDS file, and any minor errors in it become permanent as soon as the wafer begins rolling.

Wafer.space, run by Tim out of Singapore, is the shuttle that made the order possible. The designs are printed onto a single reticle on a GlobalFoundries 180 nm wafer using the GF180MCU mixed-signal process. Every time it prints, we receive a wafer with approximately a thousand dies on it after 3-4 months in the oven. Older geometry implies that the silicon is quite large by today’s standards. Unlike a Tiny Tapeout tile, this die contains only one circuit. The bundled parts we received revealed the real chip underneath, which was viewable via clear epoxy and ready to be inserted into the socket.

Bring-up was the short path, not the careful one. A carrier is slapped into a breadboard, a Raspberry Pi Pico starts up and loads a program, and the chip just counts up to 100 before shutting down. That was the proof in a single loop, with no bother. SPI failed just as expected, according to the debugging. The code for the Spade version, as well as the produced Verilog, are all publicly available, and Tong has even been distributing finished BreakingTTAPs boards to anyone who wants to experiment with one on their bench.

The second design is named BTX. That one is already in the foundry line. It’s a special processor designed specifically for ray tracing, with a completely new graphics system planned for it. Spaces on the shuttle are still available for anyone wishing to create their own chip, as you can purchase a slice of the wafer for a few thousand dollars for a thousand dies, with files expected in December 2026.

I was in the audience at Meta’s Connect conference last week when Mark Zuckerberg held up a pendant in his hands and surprise announced the Muse Charm, a little novelty-like device with Meta’s disarmingly adorable Muse AI agent onboard. Muse Charm is coming as soon as December this year.

But the funny thing is, while most people called it an AI Tamagotchi, all I saw was a strapless Apple Watch. Is that what the Muse Charm could be, eventually? A watch as well as a pendant? Is this Meta’s entry to wrist wearables in disguise? I’m still thinking about it…because smartwatches are also on their way to being AI assistants soon, too.

Reports of Meta working on its own watch wearable have been cooking for several years now. At Connect, there were smart glasses galore and even a new glasses-like VR headset coming next year. But things were suspiciously quiet on the wrist wearables front.

Meta already has one wrist wearable, the neural band that comes with the Ray-Ban Display glasses. But it received no hardware updates this year, which surprised me. The sensor-studded band uses electromyography, or EMG, to measure skin-contact electrical signals for a set of gesture controls to navigate the heads-up glasses display. Meta’s already ambitiously promised that EMG could open up all sorts of neural input possibilities, even testing the tech in a smart car dashboard prototype.

As far back as last year, Meta CTO Andrew Bosworth told me that some sort of display or watch-like evolution of the neural band could be in the works next. And Meta already introduced a partnership with Garmin to link fitness tracking to smart glasses last year.

AI agents as Meta’s new priority

The Muse Charm, however, brings a different piece of the puzzle to the table. Muse AI is Meta’s biggest priority now, clearly, based on its widespread availability on desktops and smartphones and on the company’s intentions to put the new agentic AI on all its smart glasses and, eventually, Meta’s VR Glasses. And where else could Muse work well? Enter the Charm.

The Charm has another watch-like tell. It’s powered by Qualcomm’s Snapdragon Wear Elite chipset, which already runs on Samsung’s latest smartwatches. The Wear Elite, as I learned from Qualcomm earlier this year, is optimized for a range of AI-ready wearables, ranging from watches to pendants to even glasses. There’s no reason why Meta’s Charm couldn’t easily get a wrist strap…and a watch OS to go with it.

Many reasons for a Meta smartwatch

A smartwatch would benefit Meta on several fronts. Clearly, the company wants to keep breaking into lifestyle wearables, and watches could offer another way onto people besides faces. Fitness and health tracking would give Meta data that Muse and its smart glasses could use. And watches are a key part of the glasses interface’s future. Google’s glasses will work with Wear OS watches soon. Apple’s rumored glasses will likely work with Apple Watches.

And Meta has to find a better way to make its neural band more wearable. I find the first-gen neural band fascinating in theory, but an annoying extra thing to charge in practice. If it doesn’t become a watch, or at least a fitness tracker, why wear it at all? The neural band doesn’t even control anything other than Meta’s one line of display glasses right now, but chances are Meta’s going to try to figure out ways it could work with the rest of its smart glasses lineup, and VR too.

Google and Apple are already showing how quickly smartwatches are becoming more like AI companions. Gemini lives on Wear OS watches. Apple’s adding memory-assisting audio intelligence modes to its new watches later this fall. Both can tap into Gemini and the new Siri, respectively.

Meta has Muse now. And a charm by the end of the year. Is the watch after that? Meta didn’t have any comment on the topic when I reached out to ask, but I think we’re seeing the beta test in action right in front of our eyes. Or is this the pivot to a modular idea that could be a pendant and a wristworn device, without committing to either? 

One thing’s clear, no matter what: Meta’s looking for ways for that disarmingly adorable Muse mascot to be with us all the time, even without glasses. And the Charm looks like the experiment.

Scott Stein

Editor at Large

Advertisement

I started with CNET reviewing laptops in 2009. Now I explore wearable tech, VR/AR, tablets, gaming and future/emerging trends in our changing world. Other obsessions include magic, immersive theater, puzzles, board games, cooking, improv and the New York Jets. My background includes an MFA in theater which I apply to thinking about immersive experiences of the future.

See full bio

Micron and Samsung executives this week said the memory shortage will continue for at least the next couple of years.

Micron CEO Sanjay Mehrotra expects demand for the firm’s memory to exceed its available supply over that period, he told investors last night.

Micron no longer sells consumer RAM, and Mehrotra’s statements refer to Micron’s business-to-business sales of high-bandwidth memory (HBM) for AI and DRAM for servers. However, his comments also have implications for consumer devices. Manufacturing capacity is prioritizing memory for AI and servers, limiting the supply of memory manufactured for consumer devices.

“Overall, supply-demand environment is only getting tighter,” Mehrotra said, per a transcript from Seeking Alpha.

Micron plans to open new clean rooms for memory manufacturing in 2028, but Mehrotra expects supply to remain limited.

“Even after they are built, even after first wafer output, production ramps up only gradually in the clean rooms,” he said. “That’s just the nature of what it takes to bring up production, and with HBM going from 3E to a greater mix of 4 and 4E, and with the trade ratio that exists, that … creates headwinds with respect to supply growth. Node transitions of the future give less productivity gain per wafer as well.”

Seventy-five percent of Micron’s memory output for 2027 is already accounted for, and most of the company’s current memory sales discussions are about 2028, the CEO said. Additionally, demand for HBM is surpassing demand for Micron’s DRAM.

Blaming a “significant rise” in AI submissions, Google has paused its open source bug bounty program until next year.

Last year, TechCrunch reported that cybersecurity experts were warning of that AI slop posed a serious risk to bug bounty programs. Looks like that’s the issue confronting Google’s Open Source Software Vulnerability Rewards Program, where researchers were rewarded for finding vulnerabilities in the company’s open source software.

In posts on X and the program website, Google said the bug bounty program was paused as of October 1, with a promise to provide “an update” in the first quarter of 2027. According to Tom’s Hardware, Google engineers and open source maintainers were overwhelmed by reports that were invalid or contained hallucinations.

“This pause is due to a significant rise in automated submissions, the vast majority of which are not valid,” the company said.

In the meantime, participants are encouraged to consider Google’s other bug bounty programs.

Google OSS VRP: History and Scope

Google’s Open Source Vulnerability Rewards Program (OSS VRP) was launched in August 2022. It rewards security researchers for identifying vulnerabilities in Google’s open-source software projects.

The program covers the latest versions of open-source software hosted in public repositories owned by Google on GitHub, as well as selected repositories on other platforms.

It also includes repository configuration settings, such as GitHub Actions workflows, access control rules and GitHub application configurations.

Rewards range from $100 to $31,337 based on the severity level of the reported flaws and the project’s importance.

The OSS VRP is one of several bug bounty programs operated by Google and has a relatively narrow focus.

Vulnerabilities in Google’s open source projects that are closely linked to Google Cloud or AI products are directed to the Google Cloud Vulnerability Reward Program (Cloud VRP) or the AI Vulnerability Reward Program (AI VRP), allowing reports to be routed to the teams best placed to assess and address them.

Google Plans OSS VRP Overhaul

The suspension will not affect OSS VRP supply-chain reports or any reports already submitted, Google said.

The company plans to “reformat” the program and expects to provide an update in the first quarter of 2027.

“As an alternative, we encourage you to find impact across our other VRP programs and submit there instead, or pursue the Patch Rewards Program,” Google said.

Bottom line: Steam is on course to generate more than $20 billion in gross revenue in 2026, a first for the platform, after posting record results for September and the third quarter, according to estimates from Alinea Analytics. The gains came from a mix of new releases, established franchises and free-to-play games that keep drawing spending long after launch. With one quarter left, Valve’s PC gaming platform needs roughly $3.5 billion to reach the milestone.

Steam took in an estimated $1.7 billion in September, 13% more than the previous September record, set in 2025. Its estimated third-quarter revenue reached $5.5 billion, up 12% from the same period last year.

The September result came from a familiar mix: big paid releases, durable franchises and free-to-play titles built around in-game purchases. The numbers suggest Steam no longer depends on a single kind of game sale to drive revenue.

New releases can produce a sharp burst of revenue, while games such as Counter-Strike 2 and Dota 2 keep earning money through active player communities and digital-item sales.

Through the first nine months of 2026, Steam generated an estimated $16.5 billion, compared with $14.5 billion over the same stretch of 2025. At that pace, the platform needs about $3.5 billion in the final quarter to clear $20 billion for the year.

New games played a meaningful role in September. Among Steam’s 500 top-grossing titles that month, games released in 2026 accounted for about 33.6% of revenue, according to Alinea, and new intellectual property made up 20.5% of that.

That doesn’t mean newer games displaced the industry’s biggest brands. Established franchises, which include remakes and remasters, accounted for almost 80% of revenue among the top 500. The two figures overlap because they slice the data differently, by release year and by franchise, so a 2026 sequel counts in both. Steam’s strongest performers still tend to be games with recognizable names, built-in audiences and frequent content updates.

Free-to-play games remained a major source of revenue as well. Counter-Strike 2, Apex Legends, PUBG and Dota 2 together generated almost $168 million in September, per Alinea’s estimates, or nearly 10% of Steam’s estimated monthly revenue.

Those games show why PC platforms can’t be judged by unit sales alone. Free-to-play titles may cost nothing to download, but they can drive steady spending on cosmetic items, battle passes and other digital content. For Steam, those transactions turn a large player base into recurring revenue.

Paid games accounted for several of the month’s biggest individual performances. PvP shooter Wardogs brought in an estimated $86.9 million in the roughly three weeks after it entered Early Access on September 10. Onimusha: Way of the Sword generated $30.4 million during the month, and The Blood of Dawnwalker added $26.9 million.

EA Sports FC 27 contributed another $20.7 million from Steam copies, though most of its sales were on consoles.

Wardogs’ Early Access performance is notable because it shows how large a role the model now plays on Steam. Developers can release a game before it’s finished, collect revenue early and use player feedback to shape updates. For multiplayer games, that can also help build a community before a full launch.