Third party private testing of A.I. and its source data may be essential to fair evaluation
By Wilson Miles, published by RealClear
Science
September 17, 2021 -- C-level executives
often have access to overwhelming amounts of data, yet struggle to effectively
both analyze and pull actionable insights from it. Artificial intelligence (AI)
technologies seek to address this problem by enabling computer systems—trained
on large data sets—to model human problem-solving, including finding patterns,
performing object recognition, and making predictions.
However, there are risks associated with
AI: from inaccurate facial detection causing false arrests to an AI Chatbot—designed
by the American company, OpenAI—refusing to talk about topics deemed sensitive
by the Chinese Communist Party. Perhaps the defining characteristic of
AI-enabled systems is that their decisions reflect all biases—known and
unknown—in the data with which they are trained.
This makes greater validation and
governance of automated systems, or ‘responsible AI’, critical to avoiding
AI-related accidents. Responsible AI includes striving for the maximum possible
degree of explainability and accountability. The term also refers to a
disciplined AI governance structure, in which there is active supervision in
the training and deployment of AI systems.
Responsible AI, now at the forefront of
many AI-related discussions, needs to move away from theory into practice. To
facilitate responsible AI, adopters of AI ought to pursue a certification from
an independent third party to ensure—through outside review—that the systems
they are fielding have been audited for bias which may be counter to their
operational goals.
AI systems are becoming integrated into
the real-world at an accelerating rate despite a lack of universal guidelines
on implementation and validation. As our economy, health, and security become
more dependent on these systems, their limitations create risk for companies
and customers alike. Ultimately, AI technology can put real lives at stake.
Many business leaders don’t have a clear
view into what their organization is doing to govern AI, or what new government
regulations might lie ahead. While there is effort by the European Union to
create guidelines for companies’ usage of AI, the US is lagging behind in
establishing a legal and regulatory framework to guide AI’s use, amplifying the
potential for accidental AI disasters.
As states, international organizations,
and private companies attempt to come to a consensus on regulating AI,
companies are left to navigate multiple competing viewpoints and to regulate
themselves. The question then becomes: can we trust organizations to create and
operationalize guidelines that are interpretable, fair, safe and respectful of
a user’s privacy? Even if there is a clear international understanding of
ethical AI standards, is there a mechanism for holding companies accountable to
those standards?
The best way to ensure adopters of AI
use their software responsibly is for an independent third party—from the
private sector—to certify the clients’ AI system. The certification process
entails providing best practices, which can include details on which data can
be collected and used, how models should be evaluated, and how to best deploy
and monitor models. This self-accrediting framework can also define who is
accountable for any negative outcomes of AI.
If a company's AI fails, it can likely
be attributed to the failure of either: 1) a person or people, 2) a process, or
3) the AI technology itself. An AI certification provides validation of a
company’s risk profile, inclusive of process, technology, data, people, and
culture.
A certification does pose unique
challenges. A recent GAO report stated independent audits are complicated
because an AI system can be a “black box” in which an organizations’ software
is difficult to understand, or because “vendors [will] not reveal them for proprietary
reasons.”
An AI certification should be designed
to be a self-accrediting process to avoid being a barrier to innovation, yet
simultaneously provide concrete, actionable steps—similar to the Failure Mode
and Effect Analyses approach—to identify possible issues, like unintentional
bias, and mitigate them before they cause harm to the organization.
The AI future runs on data. Both the
private and public sector need confidence in the future of AI technology if it
is to succeed en masse. Until all parties understand the benefits of AI, as
well as the ethical and national security risks posed by insufficiently
validated machine learning models, an AI certification is the only way to
achieve such assurance.
Wilson Miles is a masters candidate in
U.S. Foreign Policy and National Security at American University.
No comments:
Post a Comment