AI Memory Raises Trust Questions for OpenAI Anthropic and Microsoft

Research on AI-written personal messages finds that people can feel guilty about using AI, whether or not they have used it before. Danielle Hass of Kennesaw State University says the discomfort often reflects worries about being dishonest or manipulative. Scott Claessens notes that recipients may see AI users as lazy, less capable or less trustworthy because they value the effort behind a personal message. Harvard metaLAB’s Sarah Newman suggests asking whether you would be comfortable telling the recipient you used AI. Researchers also point to useful cases, including communicating across language barriers and helping people with dyslexia. Artist Oliva Tai’s project, presented in San Francisco, New York and Denver, collected confessions about AI use, including using it to win an argument.

In an article published Nov. 17, 2026, mEinstein founder and CEO Prithwi R. Thakuria argues that AI systems need trusted human context as they gain persistent memory and take on autonomous tasks. He warns that personal AI memory could become surveillance if others control a centralized profile. One proposed alternative is to share only specific, verified information or authorize a defined action for a particular purpose, such as a mortgage application or clinical trial. OpenAI, Anthropic and Microsoft are developing forms of persistent memory and longer-term context.

South Korea and the United States agreed to continue discussions on AI, trade, investment and cooperation in semiconductors, data centers and physical AI. At a San Francisco meeting, South Korea’s Second Vice Foreign Minister Kim Jina outlined the country’s AI Basic Act and major project plans, and sought U.S. support for Korean investment projects announced last month. The officials also discussed improving conditions for Korean businesspeople visiting the United States. Kim attended the Pax Silica Retreat and signed a declaration calling for protection of sensitive technologies and critical infrastructure, alongside growth led by super intelligence.

Databricks CEO Ali Ghodsi uses the company’s Genie AI to search information about roughly 12,000 employees, but the article stresses that AI cannot detect concerns people never put in writing. A 2003 study of 40 employees found that workers often stayed silent about a colleague’s or boss’s competence, citing fears of being labeled a troublemaker or harming a work relationship; 25% said speaking up would change nothing. Employee surveys, particularly anonymous ones, remain important because AI agents cannot replace unspoken feedback.

Atsign launched EdgeAI for organizations operating AI and connected systems across locations and networks. The package combines device and application software with the Atsign Control Centre, where administrators set communication rules based on identity rather than network location or IP address. An identity-based kill switch can block communications from compromised devices, applications or AI agents. Atsign also reports hardware-backed security work with Intel, with encryption speeds of up to 6 Gbps on Intel Core Ultra processors and 5 Gbps on Intel Xeon processors.

Several other announcements focus on deploying AI in industry and commerce. Guidance for retailers says AI shopping assistants need accurate product information—including material, style, size, color, included pieces, variants, price and availability—to match shoppers with items; retailers should not present guesses from photos or AI-generated descriptions as facts. Orange outlined AI development, alternative WAN technologies intended to provide faster, more reliable connectivity, and investments in areas including healthcare and education. SMX launched an AI-powered scientific intelligence platform, developed with a research institution, to analyze large datasets, accelerate material development and recommend manufacturing improvements that could reduce waste and costs while improving quality and productivity.

Cybersecurity analysts warn that AI agents could take a larger role in attacks by 2027, handling reconnaissance, exploitation, credential theft and movement through networks while people select targets and collect money or data. Greater automation could enable less-skilled attackers and shorten the time between initial access and harm. Organizations may also face misuse of their own AI tools, theft of credentials such as developer API keys, and “agent shadow IT”—unapproved AI tools and agents.

Key Takeaways

  • Research finds AI-written personal messages can trigger guilt and concerns about dishonesty; researchers also identify uses such as language support and assistance for people with dyslexia.
  • mEinstein founder and CEO Prithwi R. Thakuria warns that persistent AI memory needs trusted, permission-based access to human context; OpenAI, Anthropic and Microsoft are developing longer-term AI memory and context.
  • South Korea and the United States agreed to continue talks on AI, trade, investment, semiconductors, data centers and physical AI; South Korea also sought U.S. backing for investment projects announced last month.
  • Databricks CEO Ali Ghodsi uses Genie AI to search information about roughly 12,000 employees, but written data cannot reveal concerns employees keep to themselves; a cited study found 25% thought speaking up would change nothing.
  • Atsign’s EdgeAI lets administrators govern communications by identity and block compromised devices, applications or agents; Atsign reports encryption speeds up to 6 Gbps on Intel Core Ultra and 5 Gbps on Intel Xeon processors.
  • AI shopping tools depend on accurate product attributes, variant links, prices and availability; retailers should not treat image-based guesses or AI-generated descriptions as verified facts.
  • Orange outlined AI work, alternative WAN technologies and investments in healthcare and education, while SMX launched an AI platform aimed at faster material development and more efficient manufacturing.
  • AI agents could automate more stages of cyberattacks by 2027, with risks including stolen API keys, misuse of organizational AI tools and unapproved “agent shadow IT.”

AI-written personal messages can leave people feeling guilty

Researchers found that people feel guilty when they use AI to write personal messages, whether or not they have used AI before. Danielle Hass of Kennesaw State University said people often link that guilt to feeling dishonest or manipulative. Researcher Scott Claessens said AI users may be seen as lazy, less capable and less trustworthy because others value the effort behind personal messages. Sarah Newman of Harvard University’s metaLAB suggested people consider whether they would feel comfortable telling someone they used AI. Researchers said AI may help people communicate across language barriers or support people with dyslexia.

Why AI-written personal messages can make people feel guilty

Artist Oliva Tai collected confessions about AI use in a project shown in San Francisco, New York and Denver. Some people said they used AI to write personal messages, including one who used it to win an argument. Research by Danielle Hass found that people felt guilty about using AI for personal messages because they felt dishonest or manipulative. Scott Claessens said AI users may be viewed as lazier, less capable and less trustworthy. Sarah Newman suggested using discomfort as a guide, while researchers noted AI may help people communicate with a partner who speaks another language or help people with dyslexia.

mEinstein says AI needs trusted access to human context

In an article published Nov. 17, 2026, mEinstein founder and CEO Prithwi R. Thakuria argued that AI systems need trusted human context as they move toward persistent memory and autonomous tasks. The company says AI needs to know relevant details about a person while respecting what that person allows it to access or do. Thakuria warned that personal AI memory could become surveillance if it creates a centralized profile controlled by others. He described possible systems that share only specific, verified information or an authorized action for a defined purpose, such as a mortgage application or clinical trial. The article says OpenAI, Anthropic and Microsoft are developing forms of persistent memory and longer-term context for AI.

South Korea and US agree to deepen AI and investment ties

South Korea’s Second Vice Foreign Minister Kim Jina met U.S. Under Secretary of State for Economic Affairs Jacob Helberg in San Francisco on Monday. They discussed AI cooperation, trade, investment and collaboration in semiconductors, data centers and physical AI. Kim described South Korea’s AI Basic Act and major project plans, and asked for U.S. support for South Korean investment projects announced last month. The officials agreed to continue talks on economic cooperation and improve conditions for Korean businesspeople visiting the United States. Kim also attended the Pax Silica Retreat and signed the San Francisco Declaration on the Secure Foundations of Enduring Prosperity, which calls for protecting sensitive technologies and critical infrastructure and supporting growth led by super intelligence.

Databricks CEO uses AI to review staff, but it cannot hear what employees hide

Databricks CEO Ali Ghodsi uses the company’s Genie AI to find information about roughly 12,000 employees. The article notes that AI can analyze employee messages and survey responses, but cannot know about concerns employees never share in writing. A 2003 study of 40 employees found that people often kept quiet about a colleague’s or boss’s competence. Common reasons included fear of being labeled a troublemaker or damaging an important work relationship, while 25 percent said speaking up would not change anything. The article says employee surveys, especially anonymous ones, still matter, and AI agents cannot replace concerns employees choose not to express.

Atsign launches EdgeAI to secure distributed systems

Atsign launched EdgeAI for organizations running AI and connected systems across multiple locations and networks. The package combines software for devices and applications with Atsign Control Centre, where administrators can set communication rules using identity instead of network location or IP address. Its identity based kill switch can block communications from compromised devices, applications or AI agents. Atsign also described work with Intel on hardware backed security; company tests reported encryption speeds of up to 6 Gbps on Intel Core Ultra processors and 5 Gbps on Intel Xeon processors.

Accurate product data helps AI shopping tools match shoppers

AI shopping assistants rely on clear product details to match shoppers’ requests with suitable items. Retailers should provide accurate information about features such as material, style, size, color and included pieces, and avoid treating guesses from photos or AI generated descriptions as facts. Clear links between product listings and their available variants help shoppers find the exact option they want. Retailers also need to keep offer details, such as price and availability, up to date and follow each shopping service’s data requirements.

Orange highlights AI, alternative WAN and investment plans

Orange announced plans covering artificial intelligence, alternative WAN technologies and investments in sectors including healthcare and education. The company said its AI work aims to develop more efficient and effective solutions for businesses and consumers. Orange expects alternative WAN technologies to offer customers faster, more reliable connectivity. Its investments are expected to create opportunities for growth and development.

SMX launches AI platform to support American industry

SMX launched a scientific intelligence platform that uses artificial intelligence to analyze large amounts of data and find patterns. Developed through a collaboration with a research institution, the platform aims to speed up material development and help companies improve manufacturing processes. It can recommend ways to optimize production and reduce waste, which may lower costs and improve quality and productivity. The article says the platform could also create new commercial opportunities for American industry.

AI agents may carry out more cyberattacks by 2027

The article predicts that by 2027, attackers may increasingly use AI agents for reconnaissance, exploitation, credential theft and movement through networks, leaving people mainly to choose targets and collect money or data. It says most attacks still involve people making key decisions, but automated attacks could help less skilled attackers and shorten the time between first access and harm. Attackers may also misuse organizations’ own AI tools and target AI credentials, including API keys used by developers. The article expects unapproved AI tools and agents, known as “agent shadow IT,” to become a growing challenge for organizations.

Sources

NOTE:

This news brief was generated using AI technology (including, but not limited to, Google Gemini API, Llama, Grok, and Mistral) from aggregated news articles, with minimal to no human editing/review. It is provided for informational purposes only and may contain inaccuracies or biases. This is not financial, investment, or professional advice. If you have any questions or concerns, please verify all information with the linked original articles in the Sources section below.

AI ethics AI-generated personal messages AI memory Privacy and surveillance Human context OpenAI Anthropic Microsoft South Korea–U.S. AI cooperation Semiconductors Physical AI Employee feedback Databricks Genie Atsign EdgeAI Identity-based security AI in retail Industrial AI AI-powered manufacturing AI cybersecurity Autonomous AI agents

Comments

Loading...