Artificial intelligence is moving into a phase where organisations must prove not only that their systems work, but that people understand when and how those systems are being used.
New transparency obligations under the European Union’s AI Act now require disclosures around chatbots, AI agents, deepfakes and certain public-interest content. Saudi Arabia’s Communications, Space and Technology Commission has released practical guidance to help technology companies move from AI ambition to operational deployment.
Washington State University is responding to workforce demand with a new master’s programme in artificial intelligence. Telefónica, meanwhile, is warning that AI can improve immediate performance while weakening genuine learning when users begin outsourcing the thinking process itself.
Together, the developments show that AI adoption now has three inseparable requirements: transparency, organisational readiness and human capability.
EU AI transparency rules officially take effect
A major phase of the European Union’s AI Act became applicable on 2 August 2026, introducing new obligations intended to help people recognise artificial intelligence and synthetic content.
The rules respond to the growing difficulty of distinguishing authentic material from realistic images, videos, audio and text created or manipulated by AI.
According to the European Commission, the obligations are intended to reduce risks including misinformation, fraud, impersonation, manipulation and consumer deception.
The rules broadly address two situations.
The first concerns direct interaction with artificial intelligence. People must be clearly informed when they are communicating with an AI system rather than another person. This can apply to chatbots, AI agents and synthetic avatars.
The second concerns content created or substantially manipulated by AI. Providers must support the detection of certain synthetic content through machine-readable marking, while deployers may need to add visible disclosures.
Covered examples include:
- Deepfake images, audio and video that resemble real people, objects, places, organisations or events
- Emotion-recognition systems
- Biometric-categorisation tools
- AI-generated public-interest text published without human review or editorial control
- Customer-facing chatbots, agents and avatars
The distinction between a provider and deployer is important.
A provider develops or supplies the AI system and may be responsible for ensuring that its output can be technically detected. A deployer uses that system in a specific setting and may be responsible for informing the audience.
For a media company, for example, the model provider may attach provenance information to synthetic audio. The publisher using that audio may still need to tell listeners clearly that the speaker is artificial.
The Commission has also created optional icons that organisations can use when labelling AI-generated material.
Enforcement will involve national market-surveillance authorities, the European AI Office and, where European institutions are concerned, the European Data Protection Supervisor.
Companies can face penalties of up to €15 million or 3% of worldwide annual turnover. EU institutions, bodies and agencies can face fines of up to €750,000, with proportionality considerations applying to smaller businesses.
Organisations operating in Europe should consequently review every public-facing use of AI. A proper inventory should identify:
- Which AI systems interact with customers
- Where synthetic content is generated
- Whether outputs contain machine-readable marks
- Who adds visible disclosures
- Whether human editorial review occurs
- What evidence of compliance is retained
This is particularly relevant to AI-generated presenters, customer-service agents and automated media products. Labels should be understandable at the point where a person encounters the content, not hidden inside a distant policy page.
HIPTHER recently examined the approaching deadline in its AI Dispatch covering the EU AI Act, UNESCO, Chinese AI developers and global workforces.
Labelling rules introduce a dual system of visible and technical transparency
The EU’s synthetic-content requirements involve more than adding the words “created with AI” beneath an image.
Reporting from Qazinform highlights the introduction of mandatory labelling for relevant AI-generated and manipulated material.
The framework combines visible disclosure with machine-readable marking.
Visible labels are intended for people. They should be clear enough for an ordinary viewer, reader or listener to understand that material has been generated or significantly changed by AI.
Machine-readable marks are intended for platforms, detection systems and other technologies. These may include metadata, content credentials or watermarks that allow a file’s synthetic origin to be identified automatically.
Both layers are necessary.
A visible label can disappear when content is cropped or reposted. Metadata can be removed during compression or conversion. Watermarks may degrade when a file is edited, while detection tools can produce false positives and false negatives.
No single marking system provides permanent proof.
A more reliable approach combines:
- Visible labels
- Machine-readable provenance
- Platform-level detection
- Records showing which model created the content
- Information about significant edits
- Editorial review
- Procedures for correcting missing or inaccurate labels
There are also important boundaries.
Ordinary editing tools that make limited technical adjustments without substantially changing the meaning of material should not necessarily be treated in the same way as generative systems that invent realistic people, statements or events.
AI-generated text about matters of public interest requires disclosure when it has not undergone human review or editorial control. This makes human oversight legally and operationally important for news publishers.
Review must be meaningful. Passing text through an editor who merely approves it without checking claims should not be considered a reliable safeguard.
Publishers should be able to show that a human verified the material, assessed its suitability and accepted responsibility for publication.
The transparency regime is not a declaration that synthetic content is inherently harmful. AI can support accessibility, translation, entertainment and efficient communication. The objective is to ensure that its artificial origin does not become an instrument of deception.
For HIPTHER and other publishers experimenting with synthetic presenters and automated reporting, the practical lesson is straightforward: disclosures should be built into the production and delivery system rather than added manually as an afterthought.
Saudi Arabia gives technology companies an AI adoption framework
Saudi Arabia’s Communications, Space and Technology Commission has published an awareness guide designed to help technology businesses adopt AI effectively and responsibly.
The initiative coincides with the Saudi Council of Ministers designating 2026 as the Year of AI, according to TechAfrica News.
The guide is aimed at executives, strategy leaders, technology managers, product owners and departmental heads responsible for turning AI plans into operational programmes.
Its scope covers internal processes, customer-facing products and implementation models including AI agents.
The framework uses two principal perspectives.
The first examines how AI can improve a company’s internal operations and administrative processes. Potential applications include workflow automation, software development, document processing, analytics, customer support and knowledge management.
The second considers how AI can become part of the products and services offered to customers.
This distinction is useful because organisations often confuse internal productivity with external product innovation. The controls required for an employee summarisation tool differ considerably from those needed for an autonomous agent acting on behalf of customers.
The CST recommends assessing organisational readiness across five connected dimensions:
- Context readiness
- Data readiness
- Infrastructure readiness
- Skills and expertise readiness
- Organisational-culture readiness
This prevents companies from treating AI adoption as a software-purchasing exercise.
A business may have access to an advanced model but lack sufficiently reliable data. It may possess the infrastructure but not the employees needed to evaluate outputs. It may have skilled specialists while senior leadership remains unable to define appropriate use cases or risk tolerance.
Context readiness is particularly important. Companies should begin with a business problem rather than an instruction to “use AI”. A clearly defined objective makes it easier to evaluate whether automation delivers measurable value.
Data readiness requires organisations to understand what information their systems can access, whether that information is accurate and whether its use is legally permitted.
Infrastructure readiness covers compute, cloud services, APIs, cybersecurity and the systems through which AI receives information or takes action.
Skills readiness includes technical development, but also risk, legal, procurement and operational knowledge. Culture readiness determines whether employees will use the technology responsibly, challenge unreliable outputs and report incidents.
The guide reflects a wider Saudi effort to build domestic technological capability and reduce the gap between AI investment and commercial adoption.
Its strongest message is that successful deployment depends on the surrounding organisation. The model is only one component of the system.
Washington State University launches a master’s programme in AI
Washington State University has begun a new Master of Science programme in artificial intelligence as demand grows for professionals capable of developing and deploying advanced systems.
The programme is moving forward following approval by university leadership and is intended to prepare graduates for a labour market in which AI skills increasingly affect engineering, research and business roles, according to WSU News.
A dedicated AI degree can provide greater depth than adding several machine-learning modules to a conventional computer-science programme.
Students need a foundation spanning:
- Machine learning
- Data science
- Algorithm design
- Software engineering
- Statistics and mathematics
- Natural-language processing
- Computer vision
- Responsible AI
- Model evaluation
- Security and governance
Technical education must also reflect how AI is deployed in practice.
A production system requires data pipelines, permissions, monitoring, cost management and procedures for responding when a model fails. Engineers need to understand how a highly accurate model can still create harm through biased data, inappropriate automation or poor integration.
The timing is important. Employers want AI specialists, but the term can describe very different skill levels. Someone capable of operating a chatbot is not necessarily qualified to train a model, evaluate risk or build an enterprise agent.
Graduate programmes can help establish more meaningful professional standards by combining theory with hands-on engineering and research.
Universities face their own challenge, however: AI develops faster than ordinary curriculum-review cycles.
Programmes should teach durable concepts rather than focus too heavily on whichever platform currently dominates the market. Graduates need to understand model architectures, data quality, evaluation, uncertainty and system design—not merely the interface of one commercial tool.
Collaboration with employers can keep teaching connected to real applications, but universities should retain independence. Education should prepare students to assess AI critically rather than becoming product training for technology companies.
HIPTHER has previously covered attempts to strengthen the AI talent pipeline through initiatives such as IBM’s global request for AI-driven projects focused on education and the future of work.
Telefónica warns of the performance-learning paradox
AI can make a person complete a task faster while leaving them less capable of performing that task independently.
That is the central warning in a new Telefónica analysis examining how artificial intelligence is changing education and professional development.
The article describes a performance-learning paradox: AI assistance can improve immediate results while weakening the acquisition and retention of knowledge.
Telefónica cites research suggesting that generative AI can raise performance significantly during assisted work, but that the same person may perform worse when later asked to complete a task without help.
This creates what the article calls “cognitive debt”.
Like financial debt, cognitive debt offers an immediate benefit in exchange for a future cost. A user saves time by delegating analysis, writing or problem-solving, but gradually loses the ability to perform those processes independently.
Potential consequences include:
- Reduced analytical ability
- Weaker working memory
- Less original thinking
- Lower creativity
- Difficulty transferring knowledge to unfamiliar situations
- Dependence on an external system
The solution is not to ban AI from education.
Used properly, an AI system can provide personalised explanations, compare different approaches, generate exercises and act as an always-available tutor. The problem emerges when it supplies finished work before the learner has attempted to think.
Telefónica recommends a “think first” approach. Students should formulate an initial answer or strategy before asking AI for assistance. They can then use the system to challenge, compare or improve their work.
This preserves the cognitive effort needed for learning.
AI should function as a tutor rather than a substitute. A useful tutoring system asks questions, explains errors and helps learners examine alternatives. A substitution system simply produces the answer.
Assessment must also change. If institutions measure only AI-assisted output, they cannot determine whether the learner understands the subject.
Students and professionals should sometimes be evaluated without AI to distinguish tool-enabled performance from genuine knowledge. In other situations, they should be assessed on how well they use, verify and supervise AI.
This dual approach reflects the real world. People will work with AI, but they still need enough independent understanding to recognise when it is wrong.
The issue extends beyond schools and universities. Businesses risk cognitive debt when employees routinely allow AI to draft analysis, make recommendations or solve technical problems without reviewing the reasoning.
Over time, an organisation may appear more productive while losing the internal expertise needed to operate when systems fail.
HIPTHER’s coverage of OpenAI’s use of coding agents in scientific software highlighted a similar limitation: agents can execute complex implementation work, but qualified humans remain necessary to determine whether the result is scientifically valid.
The bigger picture: transparency and competence must develop together
The five developments reveal two sides of responsible AI adoption.
The EU is concentrating on transparency. People should know when they are interacting with an artificial system and when realistic content has been generated or manipulated.
Saudi Arabia’s CST is focusing on organisational readiness. Companies need suitable data, infrastructure, skills and culture before AI can deliver sustainable value.
WSU and Telefónica address human capability. Societies need more specialists who understand how to build AI, but they also need users who can think independently while working alongside it.
These goals are connected.
A visible AI label tells someone that a system was involved. It does not tell them whether the output is accurate.
An adoption framework can help a company deploy AI responsibly. It cannot compensate for employees who lack the knowledge required to challenge the system.
A specialised degree can produce qualified engineers. It cannot prevent every organisation from using AI as a shortcut around judgment.
The industry therefore needs transparency and competence to advance together.
People must be able to recognise AI, understand its limitations and retain enough independent expertise to intervene when it fails. Organisations must know what authority their systems possess and maintain evidence showing how outputs were generated, reviewed and approved.
The competitive advantage of the AI era will not belong to those who automate the greatest number of tasks. It will belong to institutions that use automation without losing accountability, knowledge or the ability to think.
