Tech
Bluesky issues its first transparency report, noting rise in user reports and legal demands
Bluesky released its first transparency report this week documenting the actions taken by its Trust & Safety team and the results of other initiatives, like age-assurance compliance, monitoring of influence operations, automated labeling, and more.
The social media startup — a rival to X and Threads — grew nearly 60% in 2025, from 25.9 million users to 41.2 million, which includes accounts hosted both on Bluesky’s own infrastructure and those running their own infrastructure as part of the decentralized social network based on Bluesky’s AT Protocol.
During the past year, users made 1.41 billion posts on the platform, which represented 61% of all posts ever made on Bluesky. Of those, 235 million posts contained media, accounting for 62% of all media posts shared on Bluesky to date.
The company also reported a fivefold increase in legal requests from law enforcement agencies, government regulators, and legal representatives in 2025, with 1,470 requests, up from 238 requests in 2024.
While the company previously shared moderation reports in 2023 and 2024, this is the first time it’s put together a comprehensive transparency report. The new report tackles other areas outside of moderation, like regulatory compliance and account verification information, among other things.
Moderation reports from users jump 54%
Compared with 2024, when Bluesky saw a 17x increase in moderation reports, the company this year reported a 54% increase, going from 6.48 million user reports in 2024 to 9.97 million in 2025.
Though the number jumped, Bluesky noted that the growth “closely tracked” its 57% user growth that occurred over the same period.
Techcrunch event
Boston, MA
|
June 23, 2026
Around 3% of the user base, or 1.24 million users, submitted reports in 2025, with the top categories being “misleading” (which includes spam) at 43.73% of the total, “harassment” at 19.93%, and sexual content at 13.54%.
A catch-all “other” category included 22.14% of the reports that didn’t fall under these categories, or others like violence, child safety, breaking site rules, or self-harm, which accounted for much smaller percentages.
Within the “misleading” category’s 4.36 million reports, spam accounted for 2.49 million reports.
Meanwhile, hate speech accounted for the largest share of the 1.99 million “harassment” reports, with about 55,400 reports. Other areas that saw activity included targeted harassment (about 42,520 reports), trolling (29,500 reports), and doxxing (about 3,170 reports).
However, Bluesky said that the majority of “harassment” reports included those that fell into the gray area of antisocial behavior, which may include rude remarks, but didn’t fit into the other categories, like hate speech.
Most of the sexual content reports (1.52 million) concerned mislabeling, Bluesky says, meaning that adult content was not properly marked with metadata — tags that allow users to control their own moderation experience using Bluesky’s tools.
A smaller number of reports focused on nonconsensual intimate imagery (about 7,520), abuse content (about 6,120), and deepfakes (over 2,000).
Reports focused on violence (24,670 in total) were broken down into subcategories like threats or incitement (about 10,170 reports), glorification of violence (6,630 reports), and extremist content (3,230 reports).
In addition to user reports, Bluesky’s automated system flagged 2.54 million potential violations.
One area where Bluesky reported success involved a decline in daily reports of antisocial behavior on the site, which dropped 79% after the implementation of a system that identified toxic replies and reduced their visibility by putting them behind an extra click, similar to what X does.
Bluesky also saw a drop in user reports month-over-month, with reports per 1,000 monthly active users declining 50.9% from January to December.
Outside of moderation, Bluesky noted it removed 3,619 accounts for suspected influence operations, most likely those operating from Russia.
Increases in takedowns, legal requests
The company said last fall it was getting more aggressive about its moderation and enforcement, and that appears to be true.
Last year, Bluesky took down 2.44 million items in 2025, including accounts and content. The year prior, Bluesky had taken down 66,308 accounts, and its automated tooling took down 35,842 accounts.
Moderators also took down 6,334 records, and automated systems removed 282.
Bluesky also issued 3,192 temporary suspensions in 2025, and 14,659 permanent removals for ban evasion. Most of the permanent suspensions were focused on accounts engaging in inauthentic behavior, spam networks, and impersonation.
However, its report suggests that it prefers labeling content more than it does booting out users. Last year, Bluesky applied 16.49 million labels to content, up 200% year-over-year, while account takedown grew 104% from 1.02 million to 2.08 million. Most of the labeling involved adult and suggestive content or nudity.
Tech
ElevenLabs CEO: Voice is the next interface for AI
ElevenLabs co-founder and CEO Mati Staniszewski says voice is becoming the next major interface for AI – the way people will increasingly interact with machines as models move beyond text and screens.
Speaking at Web Summit in Doha, Staniszewski told TechCrunch voice models like those developed by ElevenLabs have recently moved beyond simply mimicking human speech — including emotion and intonation — to working in tandem with the reasoning capabilities of large language models. The result, he argued, is a shift in how people interact with technology.
In the years ahead, he said, “hopefully all our phones will go back in our pockets, and we can immerse ourselves in the real world around us, with voice as the mechanism that controls technology.”
That vision fueled ElevenLabs’s $500 million raise this week at an $11 billion valuation, and it is increasingly shared across the AI industry. OpenAI and Google have both made voice a central focus of their next-generation models, while Apple appears to be quietly building voice-adjacent, always-on technologies through acquisitions like Q.ai. As AI spreads into wearables, cars, and other new hardware, control is becoming less about tapping screens and more about speaking, making voice a key battleground for the next phase of AI development.
Iconiq Capital general partner Seth Pierrepont echoed that view onstage at Web Summit, arguing that while screens will continue to matter for gaming and entertainment, traditional input methods like keyboards are starting to feel “outdated.”
And as AI systems become more agentic, Pierrepont said, the interaction itself will also change, with models gaining guardrails, integrations, and context needed to respond with less explicit prompting from users.
Staniszewski pointed to that agentic shift as one of the biggest changes underway. Rather than spelling out every instruction, he said future voice systems will increasingly rely on persistent memory and context built up over time, making interactions feel more natural and requiring less effort from users.
Techcrunch event
Boston, MA
|
June 23, 2026
That evolution, he added, will influence how voice models are deployed. While high-quality audio models have largely lived in the cloud, Staniszewski said ElevenLabs is working toward a hybrid approach that blends cloud and on-device processing — a move aimed at supporting new hardware, including headphones and other wearables, where voice becomes a constant companion rather than a feature you decide when to engage with.
ElevenLabs is already partnering with Meta to bring its voice technology to products, including Instagram and Horizon Worlds, the company’s virtual-reality platform. Staniszewski said he would also be open to working with Meta on its Ray-Ban smart glasses as voice-driven interfaces expand into new form factors.
But as voice becomes more persistent and embedded in everyday hardware, it opens the door to serious concerns around privacy, surveillance, and how much personal data voice-based systems will store as they move closer to users’ daily lives — something companies like Google have already been accused of abusing.
Tech
Substack confirms data breach affects users’ email addresses and phone numbers
Newsletter platform Substack has confirmed a data breach in an email to users. The company said that in October, an “unauthorized third party” accessed user data, including email addresses, phone numbers, and other unspecified “internal metadata.”
Substack specified that more sensitive data, such as credit card numbers, passwords, and other financial information, was unaffected.
In an email sent to users, Substack chief executive Chris Best said that the company identified the issue in February that allowed someone to access its systems. Best said that Substack has fixed the problem and started an investigation.
“I’m reaching out to let you know about a security incident that resulted in the email address and phone number from your Substack account being shared without your permission,” said Best in the email to users. “I’m incredibly sorry this happened. We take our responsibility to protect your data and your privacy seriously, and we came up short here.”
It’s not clear what exactly the issue was with its systems, and the scope of the data that was accessed. It’s also not yet known why the company took five months to detect the breach, or if it was contacted by hackers demanding a ransom. TechCrunch asked the company for more details, and we will update our story if we hear back.
Substack did not say how many users are affected. The company said that it doesn’t have any evidence that users’ data is being misused, but did not say what technical means, such as logs, it has to detect evidence of abuse. However, the company asked users to take caution with emails and texts without any particular indicators or direction.
On its website, Substack says that its site has more than 50 million active subscriptions, including 5 million paid subscriptions — a milestone it reached last March. In July 2025, the company raised $100 million in Series C funding led by BOND and The Chernin Group (TCG), with participation from a16z, Klutch Sports Group CEO Rich Paul, and Skims co-founder Jens Grede.
Techcrunch event
Boston, MA
|
June 23, 2026
Tech
Fundamental raises $255M Series A with a new take on big data analysis
An AI lab called Fundamental emerged from stealth on Thursday, offering a new foundation model to solve an old problem: how to draw insights from the huge quantities of structured data produced by enterprises. By combining the old systems of predictive AI with more contemporary tools, the company believes it can reshape how large enterprises analyze their data.
“While LLMs have been great at working with unstructured data, like text, audio, video, and code, they don’t work well with structured data like tables,” CEO Jeremy Fraenkel told TechCrunch. “With our model Nexus, we have built the best foundation model to handle that type of data.”
The idea has already drawn significant interest from investors. The company is emerging from stealth with $255 million in funding at a $1.2 billion valuation. The bulk of it comes from the recent $225 million Series A round led by Oak HC/FT, Valor Equity Partners, Battery Ventures, and Salesforce Ventures; Hetz Ventures also participated in the Series A, with angel funding from Perplexity CEO Aravind Srinivas, Brex co-founder Henrique Dubugras, and Datadog CEO Olivier Pomel.
Called a large tabular model (LTM) rather than a large language model (LLM), Fundamental’s Nexus breaks from contemporary AI practices in a number of significant ways. The model is deterministic — that is, it will give the same answer every time it is asked a given question — and doesn’t rely on the transformer architecture that defines models from most contemporary AI labs. Fundamental calls it a foundation model because it goes through the normal steps of pre-training and fine-tuning, but the result is something profoundly different from what a client would get when partnering with OpenAI or Anthropic.
Those differences are important because Fundamental is chasing a use case where contemporary AI models often falter. Because Transformer-based AI models can only process data that’s within their context window, they often have trouble reasoning over extremely large datasets — analyzing a spreadsheet with billions of rows, for instance. But that kind of enormous structured dataset is common within large enterprises, creating a significant opportunity for models that can handle the scale.
As Fraenkel sees it, that’s a huge opportunity for Fundamental. Using Nexus, the company can bring contemporary techniques to big data analysis, offering something more powerful and flexible than the algorithms that are currently in use.
“You can now have one model across all of your use cases, so you can now expand massively the number of use cases that you tackle,” he told TechCrunch. “And on each one of those use cases, you get better performance than what you would otherwise be able to do with an army of data scientists.”
That promise has already brought in a number of high-profile contracts, including seven-figure contracts with Fortune 100 clients. The company has also entered into a strategic partnership with AWS that will allow AWS users to deploy Nexus directly from existing instances.
Knicks acquiring G Jose Alvarado from Pelicans
Nuggets trade F Hunter Tyson to Nets, get luxury-tax relief
