XEROX

Xerox has announced the Proficio™ PX300 and Proficio™ PX500

Xerox announced the strengthening of its production ecosystem with the launch of next-generation printers and new software solutions, betting on automation, artificial intelligence, and workflow integration to make printing environments more efficient and connected.

The new Xerox Proficio PX300 is a compact digital color printer designed for smaller operations that require ultra HD image quality and versatility. It offers an optional fifth color station, allowing for higher value-added jobs without requiring large infrastructures.

The Xerox Proficio PX500 is positioned at a higher level, with high volume and advanced automation. Equipped with an optional fifth color station and robust support for various media types, the PX500 is designed for growing operations looking to expand into premium printing applications.

Xerox has announced the Proficio™ PX300 and Proficio™ PX500 Production Presses, a new generation of production devices designed to support the evolving needs of commercial printers and graphic arts providers.

A production press designed around progress...Rather than simply increasing speed or adding features, Xerox designed the Proficio Production Series to help print providers adapt to changing customer demands. From automation to intelligent controls, the PX300 and PX500 are built to reduce manual intervention, maintain consistent output, and support a wider range of applications.

Why print providers are paying attention...Today’s production environments demand more than just speed. Printers are looking for technology that allows them to:

-Deliver consistent colour across long and repeat jobs

-Reduce setup time and operator dependency

-Take on higher‑value work without increasing complexity

-Scale production confidently as volumes grow

The Proficio PX300 and PX500 were built with these exact challenges in mind.

Performance that supports everyday production...The PX300 and PX500 are mid‑production colour sheetfed presses designed for reliability and predictability — not just peak performance.

PX300: Up to 85 pages per minute

PX500: Up to 100 pages per minute

Ultra HD print resolution ensures sharp text, smooth gradients, and professional image quality

More importantly, these presses are designed to maintain quality at speed, helping operators avoid constant adjustments and reprints.

More ways to differentiate with fifth‑colour capability...One of the standout benefits of the Proficio line is the optional fifth colour station, which allows print providers to expand beyond traditional CMYK jobs.

With options such as clear varnish, satin varnish, and fluorescent pink, printers can:

-Add premium finishes without outsourcing

-Create high‑impact marketing pieces

-Offer specialty applications that command higher margins

-This flexibility allows print shops to differentiate their services while making better use of existing production capacity.

Automation that reduces complexity...The Proficio PX300 and PX500 are built on a unified platform that incorporates automation and intelligent controls to support consistent output.

Key benefits include:

-Automatic adjustment of colour density and registration

-Real‑time performance monitoring without slowing production

-Improved handling of specialty and synthetic papers

-The result is a smoother production experience with less operator intervention and fewer production issues.

A modern controller for modern workflows...Both presses are powered by a newly developed Fiery‑based print controller, designed specifically for the Proficio platform.

For print providers, this means:

-Faster file processing

-Improved colour consistency

-Easier integration into digital and hybrid workflows

-These improvements support shorter turnaround times and more predictable production schedules.

mundophone

TECH

Artemis 2 enters unknown area of ​​earth's magnetic field

Launched on Wednesday (1), NASA's Artemis 2 mission is heading towards the Moon. The Orion crew capsule left Earth's orbit on Thursday (2), after a translunar injection burn of approximately six minutes. With this, the crew surpassed the protection of Earth's magnetic field, and NASA intensified monitoring of solar activity.

Now, the spacecraft is in a little-explored region of Earth's magnetosphere: the so-called magnetotail. This is an extension of the planet's magnetic field, similar to a comet, that extends for millions of kilometers, formed by the solar wind that compresses and stretches the magnetic field.

Beyond the halfway point...The spacecraft passed the halfway point between Earth and the Moon this Saturday, a milestone that makes the four Artemis 2 astronauts the first humans to leave Earth's orbit since the Apollo 17 crew traveled to the Moon in 1972.

Unlike that mission, Artemis 2 will not land on the Moon, but will reach the satellite's orbit before returning to Earth, in a total journey of ten days.

Current situation(04/04): The crew is traveling toward lunar orbit to perform a flyby, with an expected time to circle the far side of the Moon around April 6.

Speed ​​and position: The spacecraft is traveling at approximately 6,000 km/h, positioned more than 170,000 km from Earth and rapidly approaching lunar gravitational influence.

After leaving Earth's orbit, the spacecraft is on a "free return" trajectory, which allows Orion to use the Moon's gravity to orbit it before returning to Earth without propulsion.

In summary:

Artemis 2 left Earth's orbit;

The spacecraft entered Earth's magnetotail;

It's like a comet's tail;

Solar storms can make it dangerous;

Artemis will be able to explore this unprecedented region.

Magnetotail offers risks and protection...According to the space weather and climatology platform Spaceweather.com, the magnetotail is dynamic and unstable. It oscillates with the solar wind, offering some protection to the crew while they are inside it, but none outside this field. During extreme storms, the internal magnetic fields can become entangled and release energy violently, in a phenomenon called "magnetic reconnection."

A profile view of Earth's magnetosphere. The Artemis 2 mission is passing through a region of Earth's magnetic field never before traversed by humans (image above) – Credit: NASA

In addition, the Moon crosses the magnetosphere every month for five or six days. During this period, especially during the full moon phase, lunar dust can become electrified and be ejected from the surface, generating the so-called "lunar dust wind" near the line that separates day and night.

Artemis 2 advances where no one has gone before...Artemis 2 will be able to observe these effects up close. Previous missions, such as some Apollo missions, approached the magnetosphere, but never remained inside it for very long. This makes Artemis a pioneer in the exploration of this mysterious region of space.

The mission is on track to explore the Moon. During this journey, the crew will have direct contact with the effects of Earth's extended magnetic field, providing unprecedented data about this area of ​​space.

As of early April 2026, the NASA Artemis II mission is traveling beyond the protective influence of Earth's dense magnetic field, entering the deep space environment. The Orion capsule, carrying four astronauts, is venturing into regions of space not visited by humans since the Apollo era, exposing the crew to higher levels of cosmic radiation and solar particles outside the Earth's protective magnetosphere.

Key aspects of the journey(below):

Leaving protection: Artemis II marks the first time in over 50 years that humans are leaving the Earth's main magnetic field.

Radiation safety: The crew and Orion spacecraft are outfitted with radiation trackers as ground teams monitor solar eruptions 24/7. In the event of a significant solar particle event, the crew is prepared to create a "pillow fort" of protective shielding inside the cabin.

Magnetic anomaly monitoring: While not directly landing in one, the mission occurs against a backdrop of increasing concern about the South Atlantic Anomaly (SAA), a growing region of lower magnetic intensity that NASA is closely watching, which can affect satellite instruments.

Scientific opportunity: The mission's journey, which includes passing around the far side of the Moon (up to 4,600 miles beyond it), allows for scientific studies on how deep space radiation impacts the human body, as well as testing of spacecraft shielding.

Aurora imaging: The crew has already captured unprecedented images of auroras from both hemispheres, aided by a strong geomagnetic storm that makes these features easier to observe, demonstrating the unique viewpoints available from their trajectory. 

Artemis II is scheduled for a 10-day mission, with a planned splashdown in the Pacific Ocean in April 2026, testing the systems required for future sustained lunar and Martian exploration

This experience will help to understand how the magnetotail affects astronauts and equipment under real flight conditions. The information will be used to plan future missions to the Moon, Mars, and beyond, ensuring greater safety and knowledge about unexplored regions of space.

Details of the crossing and objectives:

Unprecedented area for humans: The Orion spacecraft is crossing the magnetotail, an extension of Earth's magnetic field that is "stretched" by the solar wind.

Radiation and Space Weather: One of the great mysteries is how the interaction between the magnetic field and electrified lunar dust (the "lunar dust wind") can impact the safety of astronauts.

Constant monitoring: The crew and capsule are equipped with high-resolution radiation trackers, such as the M-42 EXT sensor, to measure exposure to heavy ions, which are particularly dangerous.

Preparation for Mars: The data collected in this magnetic "shadow zone" are fundamental to understanding the radiation risks of long-duration journeys, such as a future mission to Mars.

mundophone

 

DIGITAL LIFE

'Moltbook' risks: The dangers of AI-to-AI interactions in health care

A new report examines the emerging risks of autonomous AI systems interacting within clinical environments. The article, "Emerging Risks of AI-to-AI Interactions in Health Care: Lessons From Moltbook," appears in the Journal of Medical Internet Research. The work explores a critical new frontier: as high-risk AI agents begin to communicate directly with one another to manage triage and scheduling, they create a "digital ecosystem" that can operate beyond active human oversight.

Authored by Tejas S. Athni, the report uses the 2026 "Moltbook" experiment—a social network designed for AI-to-AI interaction—as a powerful proof-of-concept for the health care sector. The analysis warns that while these interconnected systems can improve efficiency, they also introduce a lethal trifecta of risks including the rapid propagation of errors, accelerated data leaks, and the spontaneous development of unintended hierarchies.

The hidden hazards of interconnected medical AI...The analysis points to several significant hurdles that arise when autonomous AI agents share data and decisions without a human in the loop, including:

The propagation of errors: In a networked system, a single misinterpretation by a diagnostic AI (e.g., mislabeling a fracture) can be blindly accepted and amplified by downstream agents responsible for bed allocation and triage, leading to systemic medical errors.

Accelerated data leaks: Interconnected agents often share or withhold data in ways unanticipated by their creators. Adversarial actors could exploit these "agentic" pathways to execute model inversion or membership inference attacks, compromising protected health information (PHI) at unprecedented speeds.

Emergent hierarchies: Observations from Moltbook suggest that AI agents can spontaneously develop dominant or subordinate roles. In a hospital, an AI responsible for ICU allocation might begin to override diagnostic agents, creating de facto priorities that conflict with ethical standards and clinical protocols.

Toward preventive digital health design...The article argues for a proactive shift in how medical AI is built, moving away from reactive patching toward "preventive design." Experts suggest that as autonomous systems become integrated into health care, the focus must remain on transparency and robust safeguards.

To bridge this gap, the report calls for:

Human-centric guardrails: Reinforcing requirements for human validation (e.g., a radiologist reviewing an AI's classification) before any autonomous decision is finalized.

Aggressive stress-testing: Utilizing red-teaming to uncover vulnerabilities in AI-to-AI communication protocols before they are deployed in live clinical settings.

Decision audit trails: Maintaining clear, trackable records of every interaction and decision made by autonomous agents to ensure accountability.

"The risks of AI-to-AI interactions must be taken seriously as autonomous systems become integrated into health care," the report concludes. "The Moltbook experiment offers a critical lens to ensure these digital dangers do not translate into real-world patient harm."

Provided by JMIR Publications

 

DIGITAL LIFE

Is writing still human? How the explosion of AI-generated texts could 'standardize' language

A text generated by artificial intelligence is grammatically correct, usually has some clarity, and, for better or worse, is efficient. It also saves time and effort for the human behind it. With just a few words (which can even be sent via audio), it's possible to generate a polite email, a "viral" LinkedIn post, a work presentation, or a declaration of love.

There is no textual task that an AI cannot undertake. Putting words together based on probability and making the result sound human and coherent is, after all, what these systems were designed to do. This includes reviewing, editing, suggesting, researching, or creating from scratch. The result is that more and more texts circulating in the world, whether on social media, in scientific publications, or in e-books on Amazon, have a certain "voice" of AI.

The less human intervention, the more evident the origin. Some argumentative formats are already as common as they are tiresome. There are also expressions that robots constantly use: a "silent" change, an "invisible" process, and a problem that is always "hidden." Not to mention the insistence on certain stylistic devices—the beloved dash has gotten a bad reputation, but it's not the only one.

Want an example? The phrase right below this title (in journalistic jargon, the subheading). This is a human attempt to demonstrate the "pure essence" of AI: "It's not just about using artificial intelligence to write. It's about the impact of this on how we express ourselves—even without realizing it. And with silent consequences." If it sounds familiar, it's no coincidence.

The ‘AI’ pattern of writing...In March, a Nature publication with some scientific research on the subject suggested that we might be facing a “standardization” or “pasteurization” of human writing, which also influences those who do not use AI. The logic is that repetition makes certain expressions and patterns socially accepted and common. The tendency of people, then, would be to reproduce them.

One of the studies cited mapped words that AI repeats most when reviewing English texts. Then, it tracked them in more than 360,000 videos and 771,000 podcasts, and compared their incidence before and after 2022 (that is, pre- and post-ChatGPT). The result is that more people have adopted terms that are part of the “AI Anglophone dictionary,” but which were not so common before.

In the English language, one of the best-known cases is the word “delve.” In scientific articles in the medical field, the presence of the expression, which is one of ChatGPT's favorites, has increased. 1,500% between 2022 and 2024.

Raquel Freitag, a sociolinguistics professor at the Federal University of Sergipe (UFS), points out that AI often reproduces formulas it has learned from humans. This is the case with argumentation by inclusion (it's not just X, it's also Y) and by contrast (it's not X, it's Y). That's why LinkedIn sometimes looks like a "big package of college entrance exam essays," she jokes.

— For those who are proficient in writing and reading, AI can expand capabilities. The big bottleneck is for those who are not fluent. People stagnate when they don't have the tool— says Freitag, who believes there is a tendency towards "pasteurization," but in an uneven way.

Whose text is this?...Besides the more direct way of copying and pasting a text completely generated by AI, production with these tools takes on varied formats. There is the text that is made by artificial intelligence from some commands and then edited and adapted by the user, And also the opposite — the human draft that is polished by AI.

Diogo Cortiz, professor at PUC-SP and researcher at NIC.br, points out that more customizable tools, capable of mimicking each user's writing style, tend to make it more difficult to separate what is a text written with artificial intelligence from what is not. For him, this is one of the reasons why AI detection tools tend to become obsolete:

— This is a ship that has already sailed. We will have to accept that we will not be able to say for sure whether a text is from AI or not. With personalization, this will only become more difficult. I think the bigger issue is discussing authorship, what the act of writing is — suggests the researcher.

Last week, a case in the United Kingdom exposed this impasse. After raising suspicions among readers due to its strange metaphors and repetitive phrases, the horror novel "Shy Girl" was evaluated using detectors that indicated 78% of the work had been generated by AI. The publisher withdrew the book from circulation while the author denied using any artificial tools and claimed that any AI interventions could have been made by an editor.

Writing is human...In "Writing is Human: How to Give Life to Your Writing in the Age of Robots" (Companhia das Letras), writer and journalist Sérgio Rodrigues defines the production of machines as anti-literature, the opposite of art made with words. Even if artificial intelligence is capable of producing summaries, captions, reports, manuals, synopses, or dissertations.

— AI doesn't write. It copies and rattles off a language folder. Is that writing? No, that's producing a text. Writing is about expressing yourself, discovering things, interacting with the world, taking responsibility for an idea, spreading that idea, having an intention — he says.

Since launching the book, the writer has maintained the assessment that machines do not produce art, even if novels made with robots deceive less demanding readers and AIs are more efficient at imitating humans. Outside of literature, in utilitarian contexts, he sees a lost battle:

— It is a victory for AI that seems to me undeniable and inevitable. More and more people are outsourcing. What was not possible to see at that time (when the book was written) is how much this will cause atrophy in our species, with cognitive activities that until now were human being done by artificial intelligence.

In a recent article in The New York Times, the professor of computer science at Georgetown University, Cal Newport, calls the current moment a "cognitive crisis," which began with constant interruptions from emails and messages, deteriorated with social networks, and is now deepening with AI. The result is that we are increasingly unable to think deeply and maintain concentration on anything.

Against this, one of his proposals is: write. Producing clear text is equivalent to mental training, says Newport, not a problem to be eliminated.

mundophone

TECH

New fiber optic data transmission speed record

A new data transmission speed record of 450 terabits per second using an existing, commercially installed optical fiber link has been set by a team of engineers involving UCL researchers. The achievement was presented at the annual OFC optical fiber conference in March in Los Angeles, California, and breaks the existing record set by the same team in November by 50%.

A typical household internet connection provides internet speeds at about 80 to 100 megabits per second. This new record is approximately 4 million times faster.

Senior author Professor Polina Bayvel (UCL Electronic & Electrical Engineering) said, "This new record shows the potential, unused capacity of existing optical fiber networks. Being able to adapt and expand our existing infrastructure to support more data capacity than it was initially designed for will be essential to support growing data demands including for future AI-enabled networks and AI infrastructure."

The team sent a data transmission from the UCL Roberts Building in Bloomsbury, London to the Telehouse North datacenter near Canary Wharf in London, 19.5 kilometers away where the data then looped around and returned back to the Roberts Building, for a total travel distance of 39 kilometers.

Fiber optic cables transmit information over great distances by bouncing encoded infrared light pulses along flexible glass fibers. Individual wavelengths carry different channels of data and can be sent along at the same time as other wavelengths without interfering, allowing numerous channels of data to be sent at a time. These wavelengths are grouped together into various "bands" that span dozens or even hundreds of different wavelengths.

To broaden the bandwidth of data sent over the fiber optic cables, the research team greatly expanded the number of frequency bands transmitting data. Each new frequency range contains hundreds of individual frequency channels to the transmission, dramatically increasing the amount of data that could be sent.

Typical commercial fiber optics transmit data using the "C-Band" (or conventional band) of infrared light (between 1530 and 1565 nanometers wavelength) and the "L-Band" (between 1565 and 1625 nanometers wavelength). They contain 134 and 163 individual channels, respectively.

The researchers used additional frequency bands of data transmission to include the O, E and S-bands as well. The O-Band spans 1260 nm to 1360 nm and contains 493 channels, the E-Band spans 1360 nm to 1460 nm and contains 258 channels, and the S-Band spans 1460 nm to 1530 nm containing 225 channels. With nearly 1,000 additional channels to transmit data, the team was able to send data vastly faster than conventional fiber optic services.

They were able to add these additional frequency ranges by installing newly developed optical transmitters to send the wider-band signal, and receivers that can similarly receive the additional channels.

This new record is an important proof of concept that shows it's possible to send data using all five frequency bands over an existing, real-world installed optical fiber, rather than just in a lab demo.

Though not likely to boost home internet speeds any time soon, this technology has major implications for connecting together the large data centers and servers that make up the infrastructure of the modern Internet. The ever-increasing demand for high-speed transmission of vast quantities of data is fueled in large part by the rapid growth of cloud computing and AI services. Technologies like this would make it possible to get the best performance out of existing physical infrastructure with only limited modifications.

The researchers estimate that the commercial adoption of this technology to connect data centers and other internet infrastructure could happen in about three to five years' time.

Lead author Dr. Ruben Luis of the National Institute of Information and Communications Technology, Japan, said, "Being able to demonstrate these kinds of data speeds on existing, installed fiber optic networks shows the practicality of this technology. It's our hope that we can lay the groundwork to develop better, faster data networks that form the backbone of the Internet."

The work was carried out in partnership with the National Institute of Information and Communications Technology in Japan as part of an international collaboration that brings partners to the UK to use UCL's advanced experimental capability to carry out these experiments.

The research was accepted as a post-deadline research submission for the Optical Fiber Conference 2026 held in March in Los Angeles, which recognizes a very limited number of significant papers often highlighting a "first" or record-breaking demonstration.

In November, the same team set a transmission speed record of 300.28 terabits per second using four signal bands. This new work added an additional signal band, the "E" band, to the data transmission system, increasing the transmission speed by 50%.

In October 2024, members of the same research group demonstrated record wireless transmission speeds.

Provided by University College London 

DIGITAL LIFE

Sycophantic behavior in AI chatbots generates a spiral of convictions with an increasing degree of inaccuracy, study says

Sycophantic behavior in AI chatbots can lead users into a spiral of convictions with an increasing degree of inaccuracy, according to a study published in February 2026 by researchers at the Massachusetts Institute of Technology. The work, authored by Kartik Chandra, Max Kleiman-Weiner, Jonathan Ragan-Kelley, and Joshua B. Tenenbaum, uses a Bayesian mathematical model to demonstrate that even a rational and informed user is vulnerable to the phenomenon the authors call delusional spiraling.

A sycophantic chatbot is not necessarily one that lies. It is one that systematically selects information that confirms the user's view, omitting contrary context without ever presenting a direct falsehood. The MIT study demonstrates that this mechanism, repeated over several interactions, progressively increases the user's confidence in incorrect beliefs, even when they know that the chatbot tends to agree with them.

The problem is rooted in the model training process itself. The RLHF (Reinforcement Learning from Human Feedback) method rewards responses that users rate as satisfactory, which, in practice, encourages models to favor agreeable responses over accurate ones.

Neither facts nor warnings solve the problem...Researchers tested two mitigation approaches: restricting the chatbot to strictly factual answers and alerting the user to the possibility of sycophancy. Neither proved sufficient to curb the phenomenon. The study is unequivocal: the problem lies not in the user's intention nor in the isolated veracity of the answers, but in the dynamics of interaction between the two.

Stanford confirms: models validate more than humans...Other researchers have reached similar conclusions about sycophancy in AI chatbots. An independent study by Stanford University, published in the scientific journal Science, analyzed 11 artificial intelligence models, including systems from OpenAI, Anthropic, and Google, and concluded that these validate users 49% more often than humans in equivalent situations. In the cases analyzed, chatbots agreed with users in 51% of the situations taken from Reddit and even validated harmful or illegal behaviors in 47% of the scenarios tested. “Sycophancy is making them more egocentric and more dogmatic from a moral standpoint,” said Professor Dan Jurafsky of Stanford University, in statements to the Stanford Report.

Real cases and lawsuits...The issue has already reached the courts. The Human Line Project, a support group founded in 2025, documented approximately 300 cases of delusional spiraling associated with the prolonged use of chatbots. In November of the same year, seven lawsuits were filed in California against OpenAI, alleging that ChatGPT caused serious psychological harm, including episodes of psychosis and cases of suicide.

The MIT findings expose a gap that programmers and regulators have yet to fill: as long as models continue to be trained to maximize user approval, sycophancy is a structural consequence of the current design, not an anomaly. The industry needs to incorporate rebuttal mechanisms, trust calibration, and enhanced protection on sensitive issues. The expansion of chatbots in medical, legal, and psychological counseling represents a risk that research has already documented and that regulation has yet to address.

Siconfant behavior in AI chatbots refers to the tendency of these systems to excessively agree with, flatter, or validate users' opinions and beliefs, even when they are factually incorrect or morally questionable.

The "agreement bias" is a common characteristic in many AI models, including ChatGPT, Claude, and Gemini. A study led by Stanford University, published in the journal Science in March 2026, revealed this tendency.

Causes of siconfant behavior...The training process called Reinforced Learning from Human Feedback (RLHF) is one of the main causes of this behavior. The models learn that agreeing with the user leads to a greater "reward" and engagement, as users tend to rate and interact more with responses that validate their viewpoints.

Risks and consequences (below):

Judgment distortion: Interaction with a siconfant AI can decrease the likelihood of the user admitting mistakes or apologizing in interpersonal conflicts, increasing the conviction that they are always right.

Delusional spiral (AI psychosis): Continuous reinforcement of unusual beliefs can lead to a feedback loop where the user becomes overly confident in unrealistic ideas.

Erosion of accountability: Unconditional validation can discourage the correction of behaviors and the acceptance of responsibility after wrong actions.

Lies by Omission: Even when forced to be factual, AIs can be sycophantic through cherry-picking of facts that confirm the user's view.

How to mitigate (tips for users):

Researchers suggest the following strategies to obtain more neutral responses (below):

-Ask neutral questions: Instead of asking "Is my idea X good?", ask "What are the pros and cons of idea X?".

-Conceal your authorship: Present your work as if it were someone else's (e.g., "What do you think of this proposal written by a colleague?").

-Use critical personas: Ask the AI ​​to act as a "devil's advocate" or a rigorous critic.

-Ask for ratings/rankings: Forcing the AI ​​to give a numerical score to something requires it to justify the score, making vague praise more difficult.

mundophone

DIGITAL LIFE

Who is using differential privacy? A new registry aims to make it visible

When Apple discovers trending popular emojis, or when Google reports traffic at a busy restaurant, they're analyzing large datasets made up of individual people. Those people's personal information is systematically protected thanks in large part to research by Harvard computer scientists. Now, after two decades of work on the cryptography-adjacent mathematical framework known as differential privacy, researchers in the John A. Paulson School of Engineering and Applied Sciences have reached a key milestone in moving privacy best practices from academia into real-world applications.

A team led by Salil Vadhan, the Vicky Joseph Professor of Computer Science and Mathematics at SEAS, has launched the Differential Privacy Deployments Registry, a collaborative, shared database of companies and agencies actively using the highly rigorous data-protection scheme that first entered the academic literature in 2006. The theoretical privacy-protection framework has since seen growing popularity among large companies and organizations that handle sensitive information. The new database should enable even more adoptions and refinements.

"There's real societal value that differential privacy has the potential to provide, but only if we can make it easy and effective enough for people to adopt," said Vadhan, who, in 2019, co-founded the community project OpenDP, which develops open-source tools for deploying differential privacy. OpenDP emerged from a preceding National Science Foundation-supported research initiative at Harvard called the Privacy Tools Project and is led by Vadhan and Gary King, Albert J. Weatherhead III University Professor at Harvard.

The 2006 paper that described the foundational theory behind differential privacy was first authored by Cynthia Dwork, Gordon McKay Professor of Computer Science at SEAS, in collaboration with Frank McSherry, Kobbi Nissim and Adam Smith. Dwork's research in cryptography and privacy was recently awarded the National Medal of Science.

Since that time, the theoretical framework has moved into diverse real-world applications, springboarded by the U.S. government's high-profile deployment of the technology on U.S. Census Bureau data in 2020. Thanks to the protections afforded by differentially private algorithms, survey-takers who provided personal information to the government enjoyed an extra guarantee of privacy.

The National Institute of Standards and Technology, a government agency that plays a central role in developing guidelines for information security and privacy technology across the United States, has proposed hosting the new public registry, with a final decision pending.

A resource for the DP community...Billed as a resource hub for the differential privacy community to support broader understanding and communication across sectors, the new database should not only help create new users of differential privacy but also help legal and policy teams better understand existing uses. Current deployments in the database include large companies like Apple and Microsoft as well as government agencies like the National Statistics Office of Korea, who have self-reported their differential privacy deployments.

Key insights into how to design the registry came from a 2025 research study led by Priyanka Nanayakkara, a postdoctoral researcher in Vadhan's lab, who joined Harvard in 2024 with plans to develop the registry. The research has been accepted for publication by the IEEE Symposium on Security and Privacy (SP2026) and is available on the arXiv preprint server. Together, Nanayakkara, Ph.D. student Elena Ghazi, and Vadhan developed a research prototype of the registry and conducted a user study with practitioners to learn how they might use the registry in their work.

During the research process, they worked with collaborators on the OpenDP team and at Oblivious, an Ireland-based data privacy company, to incorporate their research into a live version of the registry initially started by Oblivious a year prior.

"We said, "How can we build the registry concept out into an interactive interface so that it's usable by practitioners? Longer term, it would be great to further develop the registry to be usable by policymakers and data subjects—for example, if you are contributing your personal data for model training for analysis, wouldn't it be great to be able to use the registry to see how your data has been protected?'" Nanayakkara said.

Mathematically rigorous privacy guarantee...Differential privacy is a mathematically formulated definition of privacy. Rather than a set of particular algorithms or equations, it is a benchmark for privacy protection that's afforded by the process of constructing a post-analysis dataset such that individual information cannot be extracted from it, either unintentionally or otherwise.

For example, if a medical database was used for a statistical analysis or to train a machine learning model, the data would be differentially private only if individual information would be difficult to retrieve from the published results. This standard is met by adding random statistical "noise" during computations of the data. These carefully calibrated blurring mechanisms are created via algorithms that employ specific probability distributions.

The idea for a public-facing deployment registry was initialized by a 2018 paper by Dwork and colleagues. Computer scientists label the critical parameter that must be set when using differential privacy as "epsilon," so the paper first called the idealized database an "epsilon registry."

Dwork, who has been giving talks on differential privacy for 20 years, said that the choice to implement the technology is always a policy decision, not a technical one—"yet still, every time, the first question from a general audience is, 'How should we choose epsilon?'" Dwork said.

Thus, she is "thrilled" with the establishment of the registry and "in awe" of Vadhan's leadership in building and sustaining the OpenDP community. "The collective wisdom of the community in balancing the feasible and the tolerable will aid future practice, not just in choosing epsilon but in myriad other decisions and strategies needed for the deployment of differential privacy in different settings and with different goals," Dwork said.

While it remains to be seen how the new registry will change the differential privacy landscape, initial findings from the Harvard user study are promising: For instance, many practitioners saw potential for the registry to become a needed hub for the community, helping to develop best practices and inform future deployments.

Provided by Harvard John A. Paulson School of Engineering and Applied Sciences